Design Book › RL configs that ship

RL configs that ship

These are working examples, not laws. Each number belongs to one task at one repo version. Isaac Lab itself says to treat a shipped config as evidence for a similar workload, not as a default to copy [1].

What a designer must do

  1. Start from the config closest to your task, then write down every value you keep and why [1].
  2. Print the policy period, dt times decimation, and compare it with the hardware loop [2].
  3. Recompute samples per update when you change the environment count [3], [4].
  4. Sweep action scale for your task and rate. A copied scale gets its own sweep [5], [6].
  5. Derive every randomization range from your own uncertainty. None of these ranges came from your arm [7].
  6. Retune solver values when you change backend. They do not transfer [8].
  7. Record the repo commit with each run. The Isaac Lab v3.0.0 README warns of breaking changes, errors and regressions on this branch [9].

Scope: the shipped RL config guidelines, read from pinned repos, plus the guidelines linked in each row. All shipped-config guidelines have the level guide. The method is in RL training. Results are in Sim-to-real metrics.

How to read the tables. A dash means the database holds no row for that item. It does not mean the library default applies, and the shipped file may still set it. "Computed" means this book worked the value from the linked guidelines. "Code comment" means the file states it. Most values carry no unit in the guideline, so none is added here. (note) marks an agent-written note; see chapter 25.

The configs

The configs
Short nameTaskRepo and versionTrainerSources
ReachFranka reachIsaac Lab v3.0.0, coreRSL-RL PPO[10] to [11]
LiftFranka liftIsaac Lab v3.0.0, coreRSL-RL PPO[12] to [13]
VelocityBase velocity locomotion configIsaac Lab v3.0.0, coreEnvironment config[14] to [15]
G1Unitree G1 velocity, rough and flat runnersIsaac Lab v3.0.0, coreRSL-RL PPO[16] to [17]
GearUR10e gear assemblyIsaac Lab v3.0.0, contrib deployRSL-RL RNN PPO[18] to [19]
Typing, LabSO-101 IsaacContrib-Keyboard-SO101Isaac Lab v3.0.0, contribRSL-RL PPO[20] to [21]
Typing, NVIDIASO-101 so101-typing-taskCommit 7aec2ed, on patched Isaac Lab 3.0.0PPO[22] to [23]
Pick, stateSO-101 PickCubeMuJoCo Playground v1.0.0RSL-RL[24] to [25]
Pick, visionSO-101 SO101PickCubeVisionMuJoCo Playground v1.0.0RSL-RL[24] to [25]
StackSO-101 IsaacContrib-Stack-Cube-SO101-v0Isaac Lab v3.0.0, contribNone. Env config only, no RL runner[26] to [27]
WorkshopSO-101 Sim-to-Real WorkshopCommit ce807d9, on the Isaac Lab 2.3.2 imageNone. Imitation learning teleop, not RL[28] to [29]
RSL-RL defaultPPO constructor defaultsRSL-RL v5.5.2Library default[30] to [31]

Do not read Stack or Workshop values as RL settings. The Workshop also sits on an older major version than the rest of this catalog [28].

Environments and timing

The policy period is dt times decimation [2]. Substeps do not change it [32].

Environments and timing
ConfigEnvironmentsPhysics dtDecimationControl rateEpisodeSources
Reach40961/120430 Hz, computed12.0 s[33], [34]
Lift40961/120430 Hz, code comment-[35], [36], [37]
Velocity40960.005450 Hz, computed20.0 s[14], [38]
Gear10241.0/120.0430 Hz, computed6.66 s[39], [40], [19]
Typing, Lab40960.01 s425 Hz, code comment6.0 s[41], [42], [43], [44]
Typing, NVIDIA4096, in its measured run--25 Hz, vision policy, commands every 40 ms-[45], [46], [23]
Pick, state8192 nworld----[24]
Pick, vision4096 nworld--10 Hz, control_timestep 0.1 s-[47], [48]
Workshop, teleop-1/120260 Hz, computed-[49], [50]

No guideline states G1 timing or environment count, or says that G1 keeps the Velocity values. Other published rates, for contrast: Squint trained the SO-101 at 10 Hz and deployed at 30 Hz [51], [52], and 50 Hz is one published legged reference [53].

PPO: rollout and budget

Samples per update are environments times steps per environment [4]. The rollout time is steps times the policy period. The Pick rows treat nworld as the environment count. No guideline defines it.

PPO: rollout and budget
Confignum_steps_per_envmax_iterationssave_intervalSamples per update, computedRollout, computedSources
Reach24Preset: default 1000, hold 30005098,3040.8 s[10]
Lift3215000250131,0721.07 s[12]
G1, rough245000---[16]
G1, flat-1500---[54]
Gear5121500-524,28817.07 s[18]
Typing, Lab3215000250131,0721.28 s[20], [55]
Typing, NVIDIA3215000250131,072 at 4096 environments-[22], [45]
Pick, state241000-196,608-[24]
Pick, vision163000-65,5361.6 s[47]

max_iterations is a ceiling in a file, not a budget you must spend. The typing project's demonstrated budget is 500 P1A plus 2,000 Transit15 updates with 4,096 environments, and it is not a guaranteed minimum [56]. Evaluate between stages [57].

The steps-per-environment guideline warns that ANYmal training struggled below 25 steps or 0.5 s [58]. Reach, G1 and Pick, state ship at 24 steps [10], [16], [24], and Pick, vision at 16 [47]. Where a rollout time is computed, those tasks run longer than 0.5 s: 0.8 s and 1.6 s. So compare time as well as steps, and read that figure as ANYmal's.

PPO: learning settings

PPO: learning settings
Configlearning_rateSchedulegammalamentropy_coefdesired_klSources
RSL-RL default0.001adaptive0.990.950.010.01[59], [60], [61], [62], [31]
Reach1.0e-3adaptive0.990.950.01-[63], [64], [65]
Lift1.0e-3-0.9950.900.005-[66], [67]
G1, rough1.0e-3---0.008-[17]
Gear5.0e-4adaptive0.990.950.00.008[68], [69]
Typing, Lab1.0e-4adaptive0.9950.900.0050.01[70], [71], [72], [73], [74]
Typing, NVIDIA1.0e-4adaptive0.9950.90--[75]
Pick, state1e-3---0.005-[76]
Pick, vision1e-3-----[77]

PPO: update settings

PPO: update settings
Confignum_learning_epochsnum_mini_batchesclip_paramvalue_loss_coefmax_grad_normSources
RSL-RL default540.21.01.0[30], [78], [79], [80], [81]
Reach84---[82]
Lift54---[66]
G1, rough54---[17]
Gear816---[68]
Typing, Lab540.21.01.0[83], [84], [85], [86]
Pick, state84---[76]
Pick, vision44---[77]

Networks

Networks
Confighidden_dimsActivationObservation normalizationOtherSources
Reach[64, 64]eluOff-[87]
Lift[512, 256, 128]-On-[88]
Gear[256, 128, 64]eluOnRNN PPO[89]
Typing, Lab[512, 256, 128]eluOnShared encoder, hidden [256], latent 64. init_std 1.0[90], [91]
Typing, NVIDIA[256, 256, 128]--Transit15 sizes: observation 22, action 5. Vision policy takes 160 by 120 RGB at 25 Hz[92], [93], [46]
Pick, vision(256, 256, 128)--CNN channels (16, 32), kernels (5, 3), strides (2, 2). Wrist camera 64 by 64[94], [95], [96]

Action scale

Never copy a scale without a sweep [5]. Apply the same scale and clipping at deployment [97].

Action scale
ConfigAction typeScaleSources
Typing, LabRelative joint position, all joints0.02[98]
Typing, NVIDIAJoint-rate actions integrated into position commands, vision policy-[23]
Pick, visionDelta. Smoothing 0.50.05 on arm joints, 0.25 on gripper[99]
Pick, stateScale from force range over stiffness12.5 * FORCERANGE / stiffness on arm, 25.0 * FORCERANGE / stiffness on gripper[100], [101]
StackAbsolute joint position, five arm joints, default offset. Binary gripper, open 1.745, close 0.00.5[26], [102]
Gearjoint_action_scale0.025[103], [6]
VelocityJoint position, default offset0.5[104]
Squint, a paper, not a shipped fileSO-101 deployment scale0.15[52]

Reward terms and weights

These are only the terms the database holds. The task files may hold more, so do not read a missing term as absent.

Reward terms and weights
TaskTermWeightParametersSource
Typing, LabTyping progress2.0-[105]
Typing, LabSuccess50.0-[106]
Typing, LabMechanical power-0.0005-[107]
Typing, LabEarly termination, abnormal robot-10-[108]
Pick, SO-101Grasp1.5Open 0.80, closed 0.15, reaching std 0.025[109]
Pick, SO-101Success-Distance threshold 0.05[110]
Pick, SO-101Action rate-0.02-[111]
Pick, SO-101Floor contact-1.0-[111]
Pick, SO-101Joint velocity hinge-0.015Max velocity 0.4[112]
ReachSuccess10.0-[113]
ReachAction rate-0.0001, then -0.005 at 4500 stepsCurriculum[114], [115]
ReachAction magnitude-0.005-[114]
LiftAction L2-0.01-[116]
LiftFinger to object distance0.05std 0.4[117]
LiftEarly termination, abnormal robot-50.0-[118]
GearKeypoint tracking-1.5keypoint_scale 0.15[119]
VelocityLinear velocity tracking1.0std is math.sqrt(0.25)[120]
VelocityVertical velocity-2.0-[121]
VelocityJoint torque-1.0e-5-[122]
VelocityJoint acceleration-2.5e-7-[122]
VelocityAction rate-0.01-[15]

Typing, Lab also ends an episode on robot contact force above a threshold of 20.0 [123]. The guideline states no unit. For reward method, see the reward guidelines ([124] to [125]).

Randomization ranges

Compare these with what actually changes on your robot [126]. They are the terms the database holds, not full lists.

Randomization ranges
TaskItemRangeSource
Typing, LabKeyboard z, yaw, on reset[0.015, 0.05], [-0.1, 0.1][127]
Typing, LabKeyboard roll, on reset[0.0, 0.75][128]
Typing, NVIDIAKeyboard x, y-0.003 to 0.003 m[129]
Typing, NVIDIAKeyboard yaw-1.0 to 1.0 deg[130]
Typing, NVIDIAReset joint noise0.005 rad[130]
Pick, SO-101Cube x, y(0.20, 0.35), (-0.1, 0.1)[131]
Pick, SO-101Target cube half-extents, replace(0.012, 0.018)[132]
Pick, SO-101Control delay(0.000, 0.004) s[133]
Pick, SO-101Target resample4.0 to 7.0 s[134]
Pick, base taskgeom_friction coefficient 0, replace(0.3, 1.5)[135]
Pick, base taskActuator kp and kd, scale(0.8, 1.2)[136]
Pick, visionJoint position bias per scenario, for calibration error(-0.01, 0.01)[137]
Pick, visionEnd-effector position bias, noise(-0.0025, 0.0025), (-0.001, 0.001)[138]
StackCube x, y(0.15, 0.30), (-0.10, 0.10)[139]
StackCube yaw, minimum separation(-1.0, 1.0), 0.06[140]
ReachJoint reset, by scale(0.5, 1.5)[141]
ReachJoint position and velocity observation noise-0.01 to 0.01[142]
ReachTarget x, y(0.35, 0.65), (-0.2, 0.2)[11]
LiftObject x, y on reset[-0.2, 0.2], [-0.2, 0.2][13]
GearJoint stiffness, damping, scale(0.75, 1.5), (0.3, 3.0)[143]
GearJoint friction(0.3, 0.7)[144]
GearGear shaft position noise, reset-sampled-0.005 to 0.005[145]
GearGear friction, static and dynamic, fixed0.75, 0.75[146]
VelocityRobot-body friction, at startupStatic 0.8, dynamic 0.6[147]
VelocityRobot base mass, log-uniform scale(1 / 1.25, 1.25)[148]
VelocityPush interval10.0 to 15.0 s[149]
Workshop, not RLEgo camera focal length(12.0, 15.0)[150]
Workshop, not RLLight exposure, base task(-3.0, 1.0)[151]
Workshop, not RLLight temperature, vial task(2500.0, 9500.0)[152]

Two points stand out. The two SO-101 typing configs randomize the keyboard very differently, in the ranges above. The typing policy on the Isaac Lab page uses a fixed key map with no camera, so a moved keyboard breaks it [153]. The same project also has a vision policy [46]. The gear authors report that training without proprioceptive noise transferred well on their hardware [154]. So test that on your servos before you rely on it.

Solver settings

These hold only in the solver named [8].

Solver settings
ConfigSolverSettingsSource
Typing, LabNewton, MJWarpimplicitfast, 100 iterations, 15 ls_iterations[155], [156]
Typing, LabNewton, MJWarpnjmax 600, nconmax 600[157]
Typing, LabNewton, MJWarpPyramidal cone, impratio 1.0[158]
Typing, LabNewton, MJWarpnum_substeps 2, no MuJoCo contacts[159]
Typing, LabPhysXbounce_threshold_velocity 0.01[160]
StackNewton, MJWarpnconmax 600[161]
Pick, base taskNot statednconmax 64, njmax 256[162]
LiftPhysXbounce_threshold_velocity 0.01[163]

The typing cone differs from the general MuJoCo advice of elliptic cones with impratio about 10 for manipulation [164]. It does fit the advice to avoid large impratio values with pyramidal cones [165]. This is what "working example" means: test it for your contact.

SO-101 asset values

SO-101 asset values
AssetItemValueSource
SO101_CFGGainsFrom the USD. Stiffness and damping None[166]
SO101_CFGsoft_joint_pos_limit_factor0.98[167]
SO101_CFGSelf collision, PhysX iterationsOff, 8 position, 0 velocity[168]
SO101_CFGmax_depenetration_velocity1.0[169]
SO101_HIGH_PD_CFGArm stiffness, damping, for differential IK400.0, 80.0[170]
SO101_HIGH_PD_CFGGripper stiffness, damping17.8, 0.60[171]
SO101_HIGH_PD_CFGArm effort limit, velocity limit10.0, 10.0[172]
StackBase mount z-0.03008[27]
Workshop assetSolver iterations, root32 position, 1 velocity, fixed root link[173]
Workshop assetmax_depenetration_velocity5.0[174]
Workshop vial taskRack mass0.2[29]

Workshop joint gains, for teleoperation and not from a fit that the guideline names. Every joint has effort_limit_sim 30 [175], and the guidelines state no unit.

SO-101 asset values
Joint, as the workshop names itStiffnessDampingSource
Rotation550.7[175]
Pitch300.8[176]
Elbow250.7[177]
Wrist Pitch120.5[178]
Wrist Roll70.5[179]

Fitted SO-101 values are in Sim-to-real metrics. They differ from these and from each other. Run system identification on your arm and compare, as the check on the gtc_berlin fit guideline asks [180]. Do not mix values from different fits in one model.

Playground's real deployment homes the arm over 2.0 s by cosine interpolation at each episode start [25].

What the comparison shows

Where the database is thin

Sources

Each guideline in this chapter with its source, the date it was read and the exact words of the source. Level: must means breaking it causes failure or harm, should is good practice, guide is a rule of thumb.

  1. Treat shipped task configs as evidence for a similar workload, not as defaults to copy. Level: should. raw.githubusercontent.com 2026-11-09“Use maintained task configurations as evidence for a similar workload, not as defaults to copy.”
  2. Compute the policy period as SimulationCfg.dt times env.decimation; substeps do not change it. Level: must. isaac-sim.github.io 2026-11-08“Substeps do not change the policy period”
  3. Treat a change in parallel environment count as a training change, because it changes samples per update. Level: should. isaac-sim.github.io 2026-11-08“this also changes the number of samples per training update.”
  4. Adjust parallel environments to fit GPU memory; this also changes the number of samples per training update. Level: should. raw.githubusercontent.com 2026-11-09“Adjust parallel environments to fit your GPU's memory; this also changes the number of samples per training update.”
  5. Tune action scale per task: smaller for contact-rich work, with control rate in mind. Level: should. isaac-sim.github.io 2026-11-08“Task precision requirements (smaller for contact-rich tasks)”
  6. Treat 0.025 as the action scale of the UR10e and Rizon 4s gear task, not as a general value. 0.025 action scale Level: guide. isaac-sim.github.io 2026-11-08“self.joint_action_scale = 0.025”
  7. Derive each randomization range from how uncertain the model really is; do not widen ranges without a reason. Level: should. isaac-sim.github.io 2026-11-08“Domain randomization should span plausible modeling uncertainty, not arbitrarily wide values.”
  8. A solver setting has meaning only in the solver that uses it; Isaac Lab selects a solver but does not translate values between solvers. Level: should. raw.githubusercontent.com 2026-11-09“A configuration value has meaning only within the solver that consumes it.”
  9. The v3.0.0 README warns that this active-development branch may have breaking changes, errors and regressions. Level: should. raw.githubusercontent.com 2026-11-09“Note that this branch is currently under active development and may experience breaking changes or error messages.”
  10. The Franka reach PPO runner in Isaac Lab v3.0.0 uses num_steps_per_env = 24 and saves every 50 iterations. 24 steps Level: guide. raw.githubusercontent.com 2026-11-09“num_steps_per_env = 24 max_iterations = preset(default=1000, hold=3000) save_interval = 50”
  11. The reach task base config samples targets with x in (0.35, 0.65) and y in (-0.2, 0.2). Level: guide. raw.githubusercontent.com 2026-11-09“pos_x=(0.35, 0.65), pos_y=(-0.2, 0.2),”
  12. The Franka lift PPO base runner in Isaac Lab v3.0.0 uses num_steps_per_env = 32, max_iterations = 15000 and saves every 250. 32 steps Level: guide. raw.githubusercontent.com 2026-11-09“num_steps_per_env = 32 max_iterations = 15000 save_interval = 250”
  13. The lift task base config resets the object pose with x and y in [-0.2, 0.2]. Level: guide. raw.githubusercontent.com 2026-11-09“"x": [-0.2, 0.2], "y": [-0.2, 0.2],”
  14. The base velocity config uses decimation = 4, episode_length_s = 20.0 and sim dt = 0.005. Level: guide. raw.githubusercontent.com 2026-11-09“self.decimation = 4 self.episode_length_s = 20.0 # simulation settings self.sim.dt = 0.005”
  15. The base velocity config penalizes action rate at -0.01. -0.01 Level: guide. raw.githubusercontent.com 2026-11-09“action_rate_l2 = RewTerm(func=mdp.action_rate_l2, weight=-0.01)”
  16. The G1 rough-terrain PPO runner in Isaac Lab v3.0.0 uses num_steps_per_env = 24 and max_iterations = 5000. 24 steps Level: guide. raw.githubusercontent.com 2026-11-09“num_steps_per_env = 24 max_iterations = 5000”
  17. The G1 rough-terrain PPO uses entropy_coef = 0.008, 5 learning epochs, 4 mini-batches and learning_rate = 1.0e-3. Level: guide. raw.githubusercontent.com 2026-11-09“entropy_coef=0.008, num_learning_epochs=5, num_mini_batches=4, learning_rate=1.0e-3,”
  18. The UR10e gear assembly runner in Isaac Lab v3.0.0 uses num_steps_per_env = 512 and max_iterations = 1500. 512 steps Level: guide. raw.githubusercontent.com 2026-11-09“num_steps_per_env = 512 max_iterations = 1500”
  19. The gear assembly base config uses episode_length_s = 6.66. 6.66 s Level: guide. raw.githubusercontent.com 2026-11-09“self.episode_length_s = 6.66”
  20. The SO-101 keyboard typing PPO runner in Isaac Lab v3.0.0 uses num_steps_per_env = 32. 32 steps Level: guide. raw.githubusercontent.com 2026-11-09“num_steps_per_env = 32”
  21. The SO-101 keyboard typing task resamples its command every 10.0 s and types 1 to 5 letters. Level: guide. raw.githubusercontent.com 2026-11-09“resampling_time_range=(10.0, 10.0), debug_vis=False, letter_length=(1, 5),”
  22. The NVIDIA SO-101 typing project base runner uses num_steps_per_env = 32, max_iterations = 15000 and saves every 250. Level: guide. raw.githubusercontent.com 2026-11-09“num_steps_per_env = 32 max_iterations = 15000 save_interval = 250”
  23. The vision policy in this project integrates joint-rate actions into position commands every 40 ms. Level: guide. raw.githubusercontent.com 2026-11-09“joint-position commands every 40 ms.”
  24. The MuJoCo Playground v1.0.0 state-based pick-cube RSL-RL runner uses nworld = 8192, max_iterations = 1000 and num_steps_per_env = 24. Level: guide. raw.githubusercontent.com 2026-11-09“nworld: int = 8192 max_iterations: int = 1000 num_steps_per_env: int = 24”
  25. The Playground SO-101 real deployment homes the arm over 2.0 s by cosine interpolation at each episode start. Level: guide. raw.githubusercontent.com 2026-11-09“smoothly cosine-interpolates to the training hover pose (`--homing_duration_secs=2.0`)”
  26. The SO-101 cube stack task uses absolute joint position actions on the five arm joints with scale = 0.5 and the default offset. 0.5 Level: guide. raw.githubusercontent.com 2026-11-09“scale=0.5, use_default_offset=True,”
  27. The SO-101 cube stack task seats the base at z = -0.03008. -0.03008 Level: guide. raw.githubusercontent.com 2026-11-09“_SO101_MOUNT_Z = -0.03008”
  28. The SO-101 workshop simulation image is built from Isaac Lab 2.3.2. Level: guide. raw.githubusercontent.com 2026-11-09“FROM nvcr.io/nvidia/isaac-lab:2.3.2”
  29. The workshop vial task sets rack mass to 0.2. Level: guide. raw.githubusercontent.com 2026-11-09“rack.spawn.mass_props = sim_utils.MassPropertiesCfg(mass=0.2)”
  30. The RSL-RL v5.5.2 PPO default is num_learning_epochs = 5. 5 Level: guide. raw.githubusercontent.com 2026-11-09“num_learning_epochs: int = 5,”
  31. The RSL-RL v5.5.2 PPO default uses an adaptive schedule with desired_kl = 0.01. 0.01 Level: guide. raw.githubusercontent.com 2026-11-09“schedule: str = "adaptive", desired_kl: float = 0.01,”
  32. Each MJWarp substep uses dt divided by num_substeps, and substeps do not change the policy period. Level: guide. raw.githubusercontent.com 2026-11-09“Each solver substep uses ``SimulationCfg.dt / NewtonCfg.num_substeps``.”
  33. The reach task base config uses num_envs = 4096 and env_spacing = 2.5. 4096 envs Level: guide. raw.githubusercontent.com 2026-11-09“scene: ReachSceneCfg = ReachSceneCfg(num_envs=4096, env_spacing=2.5)”
  34. The reach task base config uses decimation = 4, episode_length_s = 12.0 and sim dt = 1/120. Level: guide. raw.githubusercontent.com 2026-11-09“self.decimation = 4 self.episode_length_s = 12.0 # simulation settings self.sim.dt = 1 / 120”
  35. The lift task base config uses num_envs = 4096 and env_spacing = 3. 4096 envs Level: guide. raw.githubusercontent.com 2026-11-09“scene: SceneCfg = SceneCfg(num_envs=4096, env_spacing=3, replicate_physics=True)”
  36. The lift task base config uses decimation = 4, which its comment states as 30 Hz. 4 Level: guide. raw.githubusercontent.com 2026-11-09“self.decimation = 4 # 30 Hz”
  37. The lift task base config uses sim dt = 1/120. Level: guide. raw.githubusercontent.com 2026-11-09“self.sim.dt = 1 / 120”
  38. The base velocity config uses num_envs = 4096 and env_spacing = 2.5. 4096 envs Level: guide. raw.githubusercontent.com 2026-11-09“scene: MySceneCfg = MySceneCfg(num_envs=4096, env_spacing=2.5)”
  39. The gear assembly base config uses num_envs = 1024 and env_spacing = 2.5. 1024 envs Level: guide. raw.githubusercontent.com 2026-11-09“scene: GearAssemblySceneCfg = GearAssemblySceneCfg(num_envs=1024, env_spacing=2.5)”
  40. The gear assembly base config uses decimation = 4 and sim dt = 1.0/120.0. Level: guide. raw.githubusercontent.com 2026-11-09“self.decimation = 4 self.sim.render_interval = self.decimation self.sim.dt = 1.0 / 120.0”
  41. The SO-101 keyboard typing task ships with num_envs = 4096 and env_spacing = 1.0. 4096 envs Level: guide. raw.githubusercontent.com 2026-11-09“scene: SO101SceneCfg = SO101SceneCfg(num_envs=4096, env_spacing=1.0, replicate_physics=True)”
  42. The SO-101 keyboard typing task uses a physics dt of 0.01 s. 0.01 s Level: guide. raw.githubusercontent.com 2026-11-09“sim: SimulationCfg = SimulationCfg(physics=PhysicsCfg(), dt=0.01)”
  43. The SO-101 keyboard typing task uses decimation = 4, which gives 100 Hz simulation and 25 Hz control. 4 Level: guide. raw.githubusercontent.com 2026-11-09“self.decimation = 4 # 100 Hz sim -> 25 Hz control”
  44. The SO-101 keyboard typing task uses episode_length_s = 6.0. 6.0 s Level: guide. raw.githubusercontent.com 2026-11-09“self.episode_length_s = 6.0”
  45. The measured rate was 36.72 s per 10 updates with 4096 environments. Level: guide. raw.githubusercontent.com 2026-11-09“At the measured 36.72 seconds per 10 updates with 4096 environments”
  46. The vision policy in this project takes 160 by 120 RGB at 25 Hz. Level: guide. raw.githubusercontent.com 2026-11-09“The actor sees 160×120 RGB at 25 Hz, stacking the previous and current images”
  47. The MuJoCo Playground v1.0.0 vision pick-cube RSL-RL runner uses nworld = 4096, max_iterations = 3000 and num_steps_per_env = 16. Level: guide. raw.githubusercontent.com 2026-11-09“nworld: int = 4096 max_iterations: int = 3000 num_steps_per_env: int = 16”
  48. The Playground SO-101 vision task runs at control_timestep = 0.1 s, a 10 Hz control rate. 0.1 s Level: guide. raw.githubusercontent.com 2026-11-09“cfg.sim.control_timestep = 0.1 # 10 Hz control frequency”
  49. The workshop SO-101 teleop environment uses decimation = 2. 2 Level: guide. raw.githubusercontent.com 2026-11-09“self.decimation = 2”
  50. The workshop SO-101 teleop environment uses sim dt = 1/120. Level: guide. raw.githubusercontent.com 2026-11-09“self.sim.dt = 1 / 120”
  51. Treat 10 Hz delta joint position control as Squint's SO-101 training setting, chosen for faster training. 10 Hz Level: guide. arxiv.org 2026-11-08
  52. On a first SO-101 transfer, consider scaling actions down and raising control rate at deployment; Squint used 0.15 and 30 Hz and reports smoother paths. 0.15 action scale Level: guide. arxiv.org 2026-11-08
  53. Use 50 Hz as one published legged reference, not a universal rate; Rudin et al. ran policies at this rate. 50 Hz Level: guide. arxiv.org 2026-11-08
  54. The G1 flat-terrain runner lowers max_iterations to 1500. 1500 iterations Level: guide. raw.githubusercontent.com 2026-11-09“self.max_iterations = 1500”
  55. The SO-101 keyboard typing runner sets max_iterations = 15000 and saves every 250 iterations. 15000 iterations Level: guide. raw.githubusercontent.com 2026-11-09“max_iterations = 15000 save_interval = 250”
  56. The demonstrated starting budget is 500 P1A plus 2,000 Transit15 updates with 4,096 environments; it is not a guaranteed minimum. Level: guide. raw.githubusercontent.com 2026-11-09“Its demonstrated starting budget is 500 P1A plus 2,000 Transit15 updates with 4,096 environments.”
  57. Evaluate between training stages, and do not treat a published update count as a guaranteed budget. Level: should. isaac-sim.github.io 2026-11-08“this is not a guaranteed minimum.”
  58. Do not drop PPO steps per environment too low; ANYmal training struggled below 25 steps, or 0.5 s of simulated time. 25 steps Level: should. arxiv.org 2026-11-08
  59. The RSL-RL v5.5.2 PPO default is gamma = 0.99. 0.99 Level: guide. raw.githubusercontent.com 2026-11-09“gamma: float = 0.99,”
  60. The RSL-RL v5.5.2 PPO default is lam = 0.95. 0.95 Level: guide. raw.githubusercontent.com 2026-11-09“lam: float = 0.95,”
  61. The RSL-RL v5.5.2 PPO default is entropy_coef = 0.01. 0.01 Level: guide. raw.githubusercontent.com 2026-11-09“entropy_coef: float = 0.01,”
  62. The RSL-RL v5.5.2 PPO default is learning_rate = 0.001. 0.001 Level: guide. raw.githubusercontent.com 2026-11-09“learning_rate: float = 0.001,”
  63. The Franka reach PPO uses learning_rate = 1.0e-3 with the adaptive schedule. 1.0e-3 Level: guide. raw.githubusercontent.com 2026-11-09“learning_rate=1.0e-3, schedule="adaptive",”
  64. The Franka reach PPO uses gamma = 0.99 and lam = 0.95. Level: guide. raw.githubusercontent.com 2026-11-09“gamma=0.99, lam=0.95,”
  65. The Franka reach PPO uses entropy_coef = 0.01. 0.01 Level: guide. raw.githubusercontent.com 2026-11-09“entropy_coef=0.01,”
  66. The Franka lift PPO uses learning_rate = 1.0e-3, entropy_coef = 0.005. Level: guide. raw.githubusercontent.com 2026-11-09“entropy_coef=0.005, num_learning_epochs=5, num_mini_batches=4, learning_rate=1.0e-3,”
  67. The Franka lift PPO uses gamma = 0.995 and lam = 0.90. Level: guide. raw.githubusercontent.com 2026-11-09“gamma=0.995, lam=0.90,”
  68. The UR10e gear assembly PPO uses entropy_coef = 0.0, 8 learning epochs and 16 mini-batches. Level: guide. raw.githubusercontent.com 2026-11-09“entropy_coef=0.0, num_learning_epochs=8, num_mini_batches=16,”
  69. The UR10e gear assembly PPO uses learning_rate = 5.0e-4 and desired_kl = 0.008. Level: guide. raw.githubusercontent.com 2026-11-09“learning_rate=5.0e-4, schedule="adaptive", gamma=0.99, lam=0.95, desired_kl=0.008,”
  70. The SO-101 keyboard typing PPO uses learning_rate = 1.0e-4 with the adaptive schedule. 1.0e-4 Level: guide. raw.githubusercontent.com 2026-11-09“learning_rate=1.0e-4,”
  71. The SO-101 keyboard typing PPO uses gamma = 0.995. 0.995 Level: guide. raw.githubusercontent.com 2026-11-09“gamma=0.995,”
  72. The SO-101 keyboard typing PPO uses lam = 0.90. 0.9 Level: guide. raw.githubusercontent.com 2026-11-09“lam=0.90,”
  73. The SO-101 keyboard typing PPO uses desired_kl = 0.01. 0.01 Level: guide. raw.githubusercontent.com 2026-11-09“desired_kl=0.01,”
  74. The SO-101 keyboard typing PPO uses entropy_coef = 0.005. 0.005 Level: guide. raw.githubusercontent.com 2026-11-09“entropy_coef=0.005,”
  75. The NVIDIA SO-101 typing project PPO uses learning_rate = 1.0e-4, gamma = 0.995 and lam = 0.90. Level: guide. raw.githubusercontent.com 2026-11-09“learning_rate=1.0e-4, schedule="adaptive", gamma=0.995, lam=0.90,”
  76. The Playground state-based pick-cube PPO uses 8 learning epochs and 4 mini-batches, entropy_coef = 0.005 and learning_rate = 1e-3. Level: guide. raw.githubusercontent.com 2026-11-09“self.algorithm.entropy_coef = 0.005 self.algorithm.num_learning_epochs = 8 self.algorithm.num_mini_batches = 4 self.algorithm.learning_rate = 1e-3”
  77. The Playground vision pick-cube PPO uses 4 learning epochs, 4 mini-batches and learning_rate = 1e-3. Level: guide. raw.githubusercontent.com 2026-11-09“self.algorithm.num_learning_epochs = 4 self.algorithm.num_mini_batches = 4 self.algorithm.learning_rate = 1e-3”
  78. The RSL-RL v5.5.2 PPO default is num_mini_batches = 4. 4 Level: guide. raw.githubusercontent.com 2026-11-09“num_mini_batches: int = 4,”
  79. The RSL-RL v5.5.2 PPO default is clip_param = 0.2. 0.2 Level: guide. raw.githubusercontent.com 2026-11-09“clip_param: float = 0.2,”
  80. The RSL-RL v5.5.2 PPO default is value_loss_coef = 1.0. 1.0 Level: guide. raw.githubusercontent.com 2026-11-09“value_loss_coef: float = 1.0,”
  81. The RSL-RL v5.5.2 PPO default is max_grad_norm = 1.0. 1.0 Level: guide. raw.githubusercontent.com 2026-11-09“max_grad_norm: float = 1.0,”
  82. The Franka reach PPO uses num_learning_epochs = 8 and num_mini_batches = 4. Level: guide. raw.githubusercontent.com 2026-11-09“num_learning_epochs=8, num_mini_batches=4,”
  83. The SO-101 keyboard typing PPO uses clip_param = 0.2 with clipped value loss. 0.2 Level: guide. raw.githubusercontent.com 2026-11-09“clip_param=0.2,”
  84. The SO-101 keyboard typing PPO uses num_learning_epochs = 5. 5 Level: guide. raw.githubusercontent.com 2026-11-09“num_learning_epochs=5,”
  85. The SO-101 keyboard typing PPO uses num_mini_batches = 4. 4 Level: guide. raw.githubusercontent.com 2026-11-09“num_mini_batches=4,”
  86. The SO-101 keyboard typing PPO uses max_grad_norm = 1.0 and value_loss_coef = 1.0. 1.0 Level: guide. raw.githubusercontent.com 2026-11-09“max_grad_norm=1.0,”
  87. The Franka reach actor and critic use hidden_dims [64, 64] with observation normalization off. Level: guide. raw.githubusercontent.com 2026-11-09“hidden_dims=[64, 64], activation="elu", obs_normalization=False,”
  88. The Franka lift actor and critic use hidden_dims [512, 256, 128] with observation normalization on. Level: guide. raw.githubusercontent.com 2026-11-09“obs_normalization=True, hidden_dims=[512, 256, 128],”
  89. The UR10e gear assembly actor and critic use hidden_dims [256, 128, 64] with observation normalization on. Level: guide. raw.githubusercontent.com 2026-11-09“hidden_dims=[256, 128, 64], activation="elu", obs_normalization=True,”
  90. The SO-101 keyboard typing actor and critic use hidden_dims [512, 256, 128] with ELU activation and a shared 64-dimension encoder latent. Level: guide. raw.githubusercontent.com 2026-11-09“hidden_dims=[512, 256, 128], activation="elu", encoder_cfg={"policy": SharedEncoderMLPModelCfg.EncoderCfg(hidden_dims=[256], latent_dim=64)},”
  91. The SO-101 keyboard typing actor starts with init_std = 1.0 and observation normalization on. 1.0 Level: guide. raw.githubusercontent.com 2026-11-09“distribution_cfg=RslRlMLPModelCfg.GaussianDistributionCfg(init_std=1.0),”
  92. The NVIDIA SO-101 typing project uses hidden_dims [256, 256, 128]. Level: guide. raw.githubusercontent.com 2026-11-09“hidden_dims=[256, 256, 128],”
  93. The Transit15 stage keeps observation and action sizes of 22 and 5. Level: guide. raw.githubusercontent.com 2026-11-09“Observation/action sizes remain 22/5.”
  94. The Playground vision pick-cube actor uses hidden_dims (256, 256, 128) with a CNN on the camera. Level: guide. raw.githubusercontent.com 2026-11-09“hidden_dims=(256, 256, 128),”
  95. The Playground vision CNN uses output channels (16, 32), kernel sizes (5, 3) and strides (2, 2). Level: guide. raw.githubusercontent.com 2026-11-09“output_channels=(16, 32), kernel_size=(5, 3), stride=(2, 2),”
  96. The Playground SO-101 vision task renders the wrist camera at 64 by 64 pixels. Level: guide. raw.githubusercontent.com 2026-11-09“cam_resolution=[(64, 64)],”
  97. Apply the same post-processing to policy outputs on the real robot as in training. Level: must. raw.githubusercontent.com 2026-11-09“**Output Consistency**: Ensuring any post-processing applied to policy outputs in Isaac Lab is also applied during real-world inference”
  98. The SO-101 keyboard typing task commands relative joint position on all joints with scale = 0.02. 0.02 Level: guide. raw.githubusercontent.com 2026-11-09“action = mdp.RelativeJointPositionActionCfg(asset_name="robot", joint_names=[".*"], scale=0.02)”
  99. The Playground SO-101 vision task uses delta actions with scale 0.05 on arm joints and 0.25 on the gripper, and smoothing 0.5. Level: guide. raw.githubusercontent.com 2026-11-09“DELTA_ACTION_SCALE = (0.05, 0.05, 0.05, 0.05, 0.05, 0.25) ACTION_SMOOTHING = 0.5”
  100. The Playground SO-101 state task scales arm actions by 12.5 times the force range over stiffness and the gripper by 25.0 times. Level: guide. raw.githubusercontent.com 2026-11-09“arm_scales = [12.5 * sim_constant.FORCERANGE / s for s in stiffnesses[:5]] gripper_scale = 25.0 * sim_constant.FORCERANGE / stiffnesses[5]”
  101. The Playground v1.0.0 SO-101 model sets FORCERANGE = 30 * 9.81 / 100 from a stated STS3215 stall torque of 30 kg-cm at 12 V. Level: guide. raw.githubusercontent.com 2026-11-09“FORCERANGE = 30 * 9.81 / 100 # STS3215 stall torque: 30 kg-cm @ 12V.”
  102. The SO-101 cube stack task uses a binary gripper with open 1.745 and close 0.0. Level: guide. raw.githubusercontent.com 2026-11-09“SO101_GRIPPER_OPEN = 1.745 SO101_GRIPPER_CLOSE = 0.0”
  103. The UR10e gear assembly config sets joint_action_scale = 0.025. 0.025 Level: guide. raw.githubusercontent.com 2026-11-09“self.joint_action_scale = 0.025”
  104. The base velocity config uses joint position actions with scale = 0.5 and the default offset. 0.5 Level: guide. raw.githubusercontent.com 2026-11-09“joint_pos = mdp.JointPositionActionCfg(asset_name="robot", joint_names=[".*"], scale=0.5, use_default_offset=True)”
  105. The SO-101 keyboard typing reward gives typing progress a weight of 2.0. 2.0 Level: guide. raw.githubusercontent.com 2026-11-09“typing_progress = RewTerm(func=mdp.letter_typing_progress, weight=2.0, params={"command_name": "typing"})”
  106. The SO-101 keyboard typing reward gives success a weight of 50.0. 50.0 Level: guide. raw.githubusercontent.com 2026-11-09“success = RewTerm(func=mdp.typing_success, weight=50.0, params={"command_name": "typing"})”
  107. The SO-101 keyboard typing reward penalizes mechanical power with a weight of -0.0005. -0.0005 Level: guide. raw.githubusercontent.com 2026-11-09“mechanical_power = RewTerm(func=mdp.mechanical_power, weight=-0.0005)”
  108. The SO-101 keyboard typing reward penalizes early termination for abnormal robot states with a weight of -10. -10 Level: guide. raw.githubusercontent.com 2026-11-09“early_termination = RewTerm(func=mdp.is_terminated_term, weight=-10, params={"term_keys": ["abnormal_robot"]})”
  109. The Playground SO-101 task weights the grasp reward at 1.5 with open 0.80, closed 0.15 and reaching std 0.025. Level: guide. raw.githubusercontent.com 2026-11-09“open_pos=0.80, closed_pos=0.15, reaching_std=0.025, weight=1.5,”
  110. The Playground SO-101 task sets the success distance threshold to 0.05. 0.05 Level: guide. raw.githubusercontent.com 2026-11-09“rewards["success"].params["distance_threshold"] = 0.05”
  111. The Playground SO-101 task sets the action rate penalty to -0.02 and the floor contact penalty to -1.0. Level: guide. raw.githubusercontent.com 2026-11-09“rewards["floor_contact"].params["penalty_scale"] = -1.0 rewards["action_rate"].params["penalty_scale"] = -0.02”
  112. The Playground SO-101 task applies a joint velocity hinge at 0.4 with penalty -0.015. Level: guide. raw.githubusercontent.com 2026-11-09“rewards["joint_vel_hinge"].params["max_vel"] = 0.4 rewards["joint_vel_hinge"].params["penalty_scale"] = -0.015”
  113. The reach task base config weights success at 10.0. 10.0 Level: guide. raw.githubusercontent.com 2026-11-09“success = RewTerm(func=mdp.is_terminated_term, weight=10.0, params={"term_keys": ["success"]})”
  114. The reach task base config penalizes action rate at -0.0001 and action magnitude at -0.005. Level: guide. raw.githubusercontent.com 2026-11-09“action_rate = RewTerm(func=mdp.action_rate_l2, weight=-0.0001) action_magnitude = RewTerm(func=mdp.action_l2, weight=-0.005)”
  115. The reach task base config raises the action rate penalty to -0.005 at 4500 steps. Level: guide. raw.githubusercontent.com 2026-11-09“func=mdp.modify_reward_weight, params={"term_name": "action_rate", "weight": -0.005, "num_steps": 4500}”
  116. The lift task base config penalizes action L2 at -0.01. -0.01 Level: guide. raw.githubusercontent.com 2026-11-09“action_l2 = RewTerm(func=mdp.action_l2, weight=-0.01)”
  117. The lift task base config rewards finger-to-object distance at 0.05 with std 0.4. Level: guide. raw.githubusercontent.com 2026-11-09“fingers_to_object = RewTerm(func=mdp.object_ee_distance, params={"std": 0.4}, weight=0.05)”
  118. The lift task base config penalizes early termination for abnormal robot state at -50.0. -50.0 Level: guide. raw.githubusercontent.com 2026-11-09“early_termination = RewTerm(func=mdp.is_terminated_term, weight=-50.0, params={"term_keys": ["abnormal_robot"]})”
  119. The UR10e gear assembly base config weights keypoint tracking at -1.5 with keypoint_scale = 0.15. Level: guide. raw.githubusercontent.com 2026-11-09“func=mdp.keypoint_entity_error, weight=-1.5, params={ "asset_cfg_1": SceneEntityCfg("factory_gear_base"), "keypoint_scale": 0.15,”
  120. The base velocity config weights linear velocity tracking at 1.0. Level: guide. raw.githubusercontent.com 2026-11-09“func=mdp.track_lin_vel_xy_exp, weight=1.0, params={"command_name": "base_velocity", "std": math.sqrt(0.25)}”
  121. The base velocity config penalizes vertical velocity at -2.0. -2.0 Level: guide. raw.githubusercontent.com 2026-11-09“lin_vel_z_l2 = RewTerm(func=mdp.lin_vel_z_l2, weight=-2.0)”
  122. The base velocity config penalizes joint torque at -1.0e-5 and joint acceleration at -2.5e-7. Level: guide. raw.githubusercontent.com 2026-11-09“dof_torques_l2 = RewTerm(func=mdp.joint_torques_l2, weight=-1.0e-5) dof_acc_l2 = RewTerm(func=mdp.joint_acc_l2, weight=-2.5e-7)”
  123. The SO-101 keyboard typing task ends an episode on robot contact force above a threshold of 20.0. 20.0 Level: guide. raw.githubusercontent.com 2026-11-09“params={"sensor_cfg": SceneEntityCfg("robot_contact"), "threshold": 20.0},”
  124. In PPO, treat time-outs differently from failures, and bootstrap the critic target at time-outs. Level: must. arxiv.org 2026-11-08
  125. The Rizon 4s gear task linearly ramps its grasp reward weight from zero so the policy first learns coarse behavior; the config shows 250,000 steps, while the prose says 512k. 512k environment steps Level: guide. isaac-sim.github.io 2026-11-08“linearly ramps the reward weight from zero over the first 512k environment steps”
  126. Pick randomization terms by asking what actually changes on the real robot and around it. Level: guide. docs.nvidia.com 2026-11-08
  127. The SO-101 keyboard typing task randomizes the keyboard pose on reset with z in [0.015, 0.05] and yaw in [-0.1, 0.1]. Level: guide. raw.githubusercontent.com 2026-11-09“"z": [0.015, 0.05], "yaw": [-0.1, 0.1],”
  128. The SO-101 keyboard typing task randomizes keyboard roll on reset in [0.0, 0.75]. Level: guide. raw.githubusercontent.com 2026-11-09“"roll": [0.0, 0.75],”
  129. The NVIDIA SO-101 typing project randomizes keyboard x and y by -0.003 to 0.003 m. Level: guide. raw.githubusercontent.com 2026-11-09“"keyboard_x_m": [ -0.003, 0.003 ], "keyboard_y_m": [ -0.003, 0.003 ],”
  130. The NVIDIA SO-101 typing project randomizes keyboard yaw by -1.0 to 1.0 deg and adds 0.005 rad reset joint noise. Level: guide. raw.githubusercontent.com 2026-11-09“"keyboard_yaw_deg": [ -1.0, 1.0 ], "reset_joint_noise_rad": 0.005”
  131. The Playground SO-101 task places the cube with x in (0.20, 0.35) and y in (-0.1, 0.1). Level: guide. raw.githubusercontent.com 2026-11-09“# x: 20 cm to 35 cm in front of the robot base. (0.20, 0.35), # y: 10 cm left to 10 cm right of the centerline. (-0.1, 0.1),”
  132. The Playground SO-101 task replaces target cube half-extents with values sampled from (0.012, 0.018). Level: guide. raw.githubusercontent.com 2026-11-09“0.012, 0.018, op="replace"”
  133. The Playground SO-101 task randomizes control delay in (0.000, 0.004) s by stochastic substeps. Level: guide. raw.githubusercontent.com 2026-11-09“delay=(0.000, 0.004),”
  134. The Playground SO-101 task resamples its target every 4.0 to 7.0 s. Level: guide. raw.githubusercontent.com 2026-11-09“resetter_params["resampling_time_range"] = (4.0, 7.0)”
  135. The Playground base pick-cube task replaces geom_friction coefficient 0 with values in (0.3, 1.5). Level: guide. raw.githubusercontent.com 2026-11-09“0: noise.uniform(0.3, 1.5, op="replace"),”
  136. The Playground base pick-cube task scales actuator kp and kd by (0.8, 1.2). Level: guide. raw.githubusercontent.com 2026-11-09“kp_noise=noise.uniform(0.8, 1.2, op="scale"), kd_noise=noise.uniform(0.8, 1.2, op="scale"),”
  137. The Playground SO-101 vision task adds a per-scenario joint-position bias in (-0.01, 0.01) to represent calibration error. Level: guide. raw.githubusercontent.com 2026-11-09“dist=noise.Uniform(low=-0.01, high=0.01),”
  138. The Playground SO-101 vision task adds an end-effector position bias in (-0.0025, 0.0025) and noise in (-0.001, 0.001). Level: guide. raw.githubusercontent.com 2026-11-09“dist=noise.Uniform(low=-0.0025, high=0.0025), op="add", ), noise=noise.Noise( dist=noise.Uniform(low=-0.001, high=0.001),”
  139. The SO-101 cube stack task places cubes on reset with x in (0.15, 0.30) and y in (-0.10, 0.10). Level: guide. raw.githubusercontent.com 2026-11-09“"x": (0.15, 0.30), "y": (-0.10, 0.10),”
  140. The SO-101 cube stack task randomizes cube yaw in (-1.0, 1.0) with a minimum separation of 0.06. Level: guide. raw.githubusercontent.com 2026-11-09“"yaw": (-1.0, 1.0), } self.events.randomize_cube_positions.params["min_separation"] = 0.06”
  141. The reach task base config resets joints by scale in (0.5, 1.5). Level: guide. raw.githubusercontent.com 2026-11-09“"position_range": (0.5, 1.5),”
  142. The reach task base config adds uniform noise of -0.01 to 0.01 to joint position and joint velocity observations. Level: guide. raw.githubusercontent.com 2026-11-09“joint_pos = ObsTerm(func=mdp.joint_pos_rel, noise=Unoise(n_min=-0.01, n_max=0.01))”
  143. The UR10e gear assembly base config scales joint stiffness by (0.75, 1.5) and joint damping by (0.3, 3.0). Level: guide. raw.githubusercontent.com 2026-11-09“"stiffness_distribution_params": (0.75, 1.5), "damping_distribution_params": (0.3, 3.0), "operation": "scale",”
  144. The UR10e gear assembly config adds joint friction from a distribution of (0.3, 0.7). Level: guide. raw.githubusercontent.com 2026-11-09“"friction_distribution_params": (0.3, 0.7),”
  145. The UR10e gear assembly base config adds reset-sampled position noise of -0.005 to 0.005 to the gear shaft position observation. Level: guide. raw.githubusercontent.com 2026-11-09“noise_cfg=UniformNoiseCfg(n_min=-0.005, n_max=0.005, operation="add")”
  146. The UR10e gear assembly base config sets gear friction to static 0.75 and dynamic 0.75. Level: guide. raw.githubusercontent.com 2026-11-09“"static_friction_range": (0.75, 0.75), "dynamic_friction_range": (0.75, 0.75),”
  147. The base velocity config sets robot-body static friction to 0.8 and dynamic friction to 0.6 at startup. Level: guide. raw.githubusercontent.com 2026-11-09“"static_friction_range": (0.8, 0.8), "dynamic_friction_range": (0.6, 0.6),”
  148. The base velocity config scales robot base mass using a log-uniform distribution of (1 / 1.25, 1.25). Level: guide. raw.githubusercontent.com 2026-11-09“"mass_distribution_params": (1 / 1.25, 1.25), "operation": "scale",”
  149. The base velocity config applies pushes at an interval of 10.0 to 15.0 s. Level: guide. raw.githubusercontent.com 2026-11-09“interval_range_s=(10.0, 15.0),”
  150. The workshop randomizes the ego camera's focal length in (12.0, 15.0). Level: guide. raw.githubusercontent.com 2026-11-09“"focal_length_range": (12.0, 15.0),”
  151. The workshop randomizes light exposure in (-3.0, 1.0) in the base task. Level: guide. raw.githubusercontent.com 2026-11-09“"exposure_range": (-3.0, 1.0),”
  152. The workshop vial task randomizes light temperature in (2500.0, 9500.0). Level: guide. raw.githubusercontent.com 2026-11-09“"temperature_range": (2500.0, 9500.0),”
  153. The typing policy uses a fixed key map with no camera, so a moved keyboard breaks it. Level: must. raw.githubusercontent.com 2026-11-09“The policy uses a fixed A-Z key-position map without camera input; it cannot adjust to a moved keyboard.”
  154. The gear assembly authors report that training without proprioceptive noise transferred well on their hardware. Level: guide. raw.githubusercontent.com 2026-11-09“Empirically, we found that policies trained without noise on proprioceptive observations (joint positions and velocities) transfer well to real hardware.”
  155. The SO-101 keyboard typing task on Newton uses the MJWarp solver with the implicitfast integrator. Level: guide. raw.githubusercontent.com 2026-11-09“integrator="implicitfast",”
  156. The SO-101 keyboard typing task on Newton sets iterations = 100 and ls_iterations = 15. 100 Level: guide. raw.githubusercontent.com 2026-11-09“iterations=100, ls_iterations=15,”
  157. The SO-101 keyboard typing task on Newton sets njmax = 600 and nconmax = 600. 600 Level: guide. raw.githubusercontent.com 2026-11-09“nconmax=600,”
  158. The SO-101 keyboard typing task on Newton uses a pyramidal friction cone with impratio = 1.0. Level: guide. raw.githubusercontent.com 2026-11-09“impratio=1.0, cone="pyramidal",”
  159. The SO-101 keyboard typing task on Newton uses num_substeps = 2 and does not use MuJoCo contacts. 2 Level: guide. raw.githubusercontent.com 2026-11-09“num_substeps=2,”
  160. The SO-101 keyboard typing task on PhysX sets bounce_threshold_velocity = 0.01. 0.01 Level: guide. raw.githubusercontent.com 2026-11-09“bounce_threshold_velocity=0.01,”
  161. The SO-101 cube stack task sets nconmax = 600 on Newton MJWarp. 600 Level: guide. raw.githubusercontent.com 2026-11-09“self.sim.physics.newton_mjwarp.solver_cfg.nconmax = 600”
  162. The Playground base pick-cube task sets nconmax = 64 and njmax = 256. Level: guide. raw.githubusercontent.com 2026-11-09“nconmax=64, njmax=256,”
  163. The lift task base config sets the PhysX bounce threshold velocity to 0.01. 0.01 Level: guide. raw.githubusercontent.com 2026-11-09“bounce_threshold_velocity=0.01,”
  164. Use elliptic cones with impratio about 10 for MuJoCo manipulation and 10-100 for legged robots when slip matters. 10 impratio Level: should. mujoco.readthedocs.io 2026-11-08
  165. Avoid large impratio values when MuJoCo uses pyramidal cones. Level: should. mujoco.readthedocs.io 2026-11-08
  166. The default SO-101 asset in Isaac Lab v3.0.0 takes actuator gains from the USD, with stiffness and damping set to None in the implicit actuator group. Level: guide. raw.githubusercontent.com 2026-11-09“stiffness=None,”
  167. The default SO-101 asset uses soft_joint_pos_limit_factor = 0.98. 0.98 Level: guide. raw.githubusercontent.com 2026-11-09“soft_joint_pos_limit_factor=0.98,”
  168. The default SO-101 asset turns self collisions off with 8 PhysX position iterations and 0 velocity iterations. Level: guide. raw.githubusercontent.com 2026-11-09“enabled_self_collisions=False, solver_position_iteration_count=8, solver_velocity_iteration_count=0”
  169. The default SO-101 asset sets max_depenetration_velocity = 1.0. 1.0 Level: guide. raw.githubusercontent.com 2026-11-09“rigid_props=PhysxRigidBodyCfg(disable_gravity=False, max_depenetration_velocity=1.0),”
  170. The high-PD SO-101 arm group uses stiffness 400.0 and damping 80.0 for differential IK. 400.0 Level: guide. raw.githubusercontent.com 2026-11-09“stiffness=400.0, damping=80.0,”
  171. The high-PD SO-101 gripper group uses stiffness 17.8 and damping 0.60. 17.8 Level: guide. raw.githubusercontent.com 2026-11-09“stiffness=17.8, damping=0.60,”
  172. The high-PD SO-101 arm group limits effort to 10.0 and velocity to 10.0. Level: guide. raw.githubusercontent.com 2026-11-09“joint_effort_limit=10.0, joint_velocity_limit=10.0,”
  173. The workshop SO-101 asset uses 32 solver position iterations, 1 solver velocity iteration and a fixed root link. Level: guide. raw.githubusercontent.com 2026-11-09“solver_position_iteration_count=32, solver_velocity_iteration_count=1, fix_root_link=True,”
  174. The workshop SO-101 asset sets max_depenetration_velocity = 5.0. 5.0 Level: guide. raw.githubusercontent.com 2026-11-09“max_depenetration_velocity=5.0,”
  175. The workshop SO-101 Rotation joint uses stiffness 55 and damping 0.7 with effort limit 30. Level: guide. raw.githubusercontent.com 2026-11-09“effort_limit_sim=30, stiffness=55, damping=0.7,”
  176. The workshop SO-101 Pitch joint uses stiffness 30 and damping 0.8. Level: guide. raw.githubusercontent.com 2026-11-09“effort_limit_sim=30, stiffness=30, damping=0.8,”
  177. The workshop SO-101 Elbow joint uses stiffness 25 and damping 0.7. Level: guide. raw.githubusercontent.com 2026-11-09“effort_limit_sim=30, stiffness=25, damping=0.7,”
  178. The workshop SO-101 Wrist Pitch joint uses stiffness 12 and damping 0.5. Level: guide. raw.githubusercontent.com 2026-11-09“effort_limit_sim=30, stiffness=12, damping=0.5,”
  179. The workshop SO-101 Wrist Roll joint uses stiffness 7 and damping 0.5. Level: guide. raw.githubusercontent.com 2026-11-09“effort_limit_sim=30, stiffness=7, damping=0.5,”
  180. The gtc_berlin fit gives the SO-101 shoulder pan stiffness 48.93682087435154, damping 2.309705389862818 and armature 0.058228342618297214. Level: guide. raw.githubusercontent.com 2026-11-09“stiffness: 48.93682087435154 damping: 2.309705389862818 armature: 0.058228342618297214”
  181. Keep self-collision off for training; it defaults to off and turning it on slows training. Level: should. raw.githubusercontent.com 2026-11-09
  182. The gear assembly page reports ±5 mm position noise for UR10e and ±10 mm for Rizon 4s to match their perception error. Level: guide. raw.githubusercontent.com 2026-11-09“The position noise is doubled (±10mm vs ±5mm)”
  183. Have agents read docs at the pinned simulator version, because the Isaac Lab 3.0.0 project notes that compatibility with later releases has not been validated. Level: must. isaac-sim.github.io 2026-11-08“or later versions has not been validated.”