Design Book › Sim-to-sim and sim-to-real

Sim-to-sim and sim-to-real

What a designer must do

  1. Freeze the policy interface and diff it across backends [1], [2], [3].
  2. Confirm the same task trains in both engines [4].
  3. Treat a successful sim-to-sim transfer as a possible first step toward sim-to-real deployment [5].
  4. Fit the model to hardware logs. Do not hand-tune it [6].
  5. Report the gap per joint in position, velocity and torque [7].
  6. Classify each gap before you pick a fix [8].
  7. Test on hardware through a written plan [9].

Scope: the sim-to-sim and sim-to-real guidelines. Treat transfer as two stages. Sim-to-sim tests that a policy does not exploit one solver [10], and a success can be a first step toward deployment [5]. Sim-to-real tests the model against your hardware [9].

Gate A: sim-to-sim interface

All rows are from the Isaac Lab 3.0.0 docs, Transfer Policies Between PhysX and Newton. Every field in the table must match exactly, or you are testing a different task.

Gate A: sim-to-sim interface
LevelMust matchCheckSource
mustActions: term order, joint names, width, target type, scale, offset, clippinglint: diff resolved action configs[1]
mustObservations: groups, term order, width, history, units, frames, clipping, corruptionlint: diff resolved observation configs[2]
mustPolicy state: normalization statistics, recurrent state shape and reset, commandsreview: bundle holds the statistics[11]
mustTiming: physics dt, decimation, policy period, action hold. Only Newton substeps may differ within a policy periodlint: diff timing fields[3]
mustMechanism: body and joint order, active joints, mimic or equality linkslint: compare ordered name lists[12]
mustEpisode: reset and command sampling, reward and termination meaning, length, success testlint: diff the episode config[13]
mustA physics preset must not silently change policy-facing terms. If it does, it is a different tasklint: diff MDP terms across presets[14]
mustBranched robots: one joint and body ordering convention for training and all playbacklint: print joint order per backend[15]

Gate B: sim-to-sim test

Gate B: sim-to-sim test
LevelRequirementCheckSource
shouldStep one identical fixed state in each backend over several resets, before training or transfersim: record object motion, contact count, gripper effort, penetration, success[16]
shouldThe same task trains in both enginessim: train once on each[4]
shouldExpect similar behaviour, not identical trajectoriessim: compare task metrics[17]
guideTreat a successful transfer as a possible first step toward sim-to-real deploymentsim: evaluate the checkpoint in a second backend[5]
shouldReproduce a Newton-only failure at its first failing step: one environment, fixed seed and reset state, no randomization, the same actions in both backendssim: store this reproduction[18]
shouldLog how often actuators saturate and action signs flipsim: both, per joint[19]
shouldDo not port solver settings by number. A value means something only within its solverreview: each backend has its own tuning record[20]

Doc for these rows: Isaac Lab 3.0.0 docs, in the Transfer Policies, Prepare an Asset and Solver Differences pages.

Failure signatures

Failure signatures
SymptomSuspect firstSource
Locomotion policy falls within a few dozen steps in every environmentJoint order, before contacts or friction[21]
Grip force differs between backends at one commandDoubled effort on PhysX, when both mimic fingers have nonzero drives and receive one logical command[22]
Position policy alternates saturated commandsIt exploits one solver's response. Raise damping[10]
Friction behaves differently after a portBackend values do not map one-to-one[23], [24]
Policy shakes violently on hardwareUnmodeled delay[25]
Small commanded motions are not reached on hardwareStiction not modeled. The policy learned that commands are always reached[26]
Motions land off target after a bench changeA moved fixture invalidates the layout check. The typing policy cannot adjust to a moved keyboard[27], [28]

Gate C: system identification

Fit the model to your robot [6]. Give each parameter physically justified bounds [29].

Gate C: system identification
LevelDoc as the guideline states itRequirementCheckSource
shouldMuJoCo 4.0.0 docs, ModelingFit armature, damping, friction, gains, inertia and delay to hardware logs with optimizationmeasure: command and sensor logs, then mujoco.sysid[6]
mustMuJoCo 4.0.0 docs, Pythonmujoco.sysid uses box-constrained least squares, so give each parameter real boundsreview: bounds are physically justified[29]
shouldMuJoCo 4.0.0 docs, PythonUse the physically consistent 10-D pseudo-inertia helper for link inertiasim: assert a valid inertia after fitting[30]
shouldMuJoCo 4.0.0 docs, PythonInclude sensor and actuator delays and biases when they existmeasure: fit with and without, compare held-out error[31]
guidemujoco.sysid README, main branch, captured 2026-11-08Budget one full rollout per parameter perturbationsim: time one rollout, multiply by parameter count[32]

Use the direct negative solref format in identification models [33]. Combine identification with randomization, not one or the other [34]. Retune damping after increasing armature [35].

Delay

Delay
LevelRequirementCheckSource
shouldMeasure control loop timing on hardware. It differs from simulationmeasure: timestamp commands, compute jitter[36]
mustMuJoCo delay at or below timestep times nsample. Beyond that it extrapolates non-causally with no runtime errorlint: on each actuator and sensor[37]
shouldThe smallest effective positive MuJoCo delay is one timestepsim: step a test signal[38]
shouldFor stochastic delay or identification, set nsample from the largest delaylint: nsample is at least the largest delay divided by the timestep, rounded up[39]
shouldFor continuous-valued quantities in delay identification, use linear or cubic interp to avoid zero-order-hold discontinuity. Not for integer-valued sensorssim: residual is continuous across a sweep[40]
guidemjlab replays delay quantized to the physics timestep, so a fine delay needs a fine stepreview: resolution against measured delay[41]

Friction

Friction
LevelRequirementCheckSource
shouldPlot a real step response to quantify stiction, then add joint friction. The UR10e showed about 0.25 degrees of offsetmeasure: steady-state offset[42]
shouldModel joint friction where stiction existssim: train with and without, compare real error[26]
guideCompare sim and real video of contact while tuning, without waiting for full trainingmeasure: same grasp on both[43]

Actuators

Actuators
LevelRequirementCheckSource
mustDo not assume an ideal actuator model transfers on series elastic drives. On ANYmal, ideal and analytical models both failedmeasure: predicted against measured torque[44]
shouldSuspect unmodeled delay when a policy shakes violentlymeasure: command-to-response delay per joint[25]
guideConsider a learned model when hand-tuned models stall. The ANYmal team spent over a week without much successmeasure: torque error per model[45]
guideActuator network: joint torque from position error and velocity historymeasure: held-out logs[46]
shouldCheck independence before one per-actuator network. Shared hydraulic supply may couple actuatorsmeasure: neighbours active and idle[47]
shouldModel friction, backlash and thermal effects where they matter. Typical motor models lack themmeasure: step responses cold and warm[48]

Gap metrics

Gap metrics
MetricHowSource
Paired logsReplay one motion file on the robot and in sim[49]
Per-joint tablePosition, velocity and torque, not one robot-wide number[7]
Error statisticsRMSE, MAPE, correlation, cosine similarity per joint[50]
Steady-state offsetStep response on the real joint[42]
DelayCommand-to-response per joint[25]
Loop timingTimestamp commands, compute jitter[36]
DimensionsKey link lengths on the built robot against CAD[51]
MassWeigh each removable link[52]
Task successA stated trial count per task, with the simulation figure beside it[53]
Simulation figureQuote only with its training-time condition[54]

SAGE computes the error statistics, but check it first. Its README badges and requirements name Isaac Lab 2.2.0 and Isaac Sim 5.0.0 [55]. The catalog's release records list Isaac Lab v3.0.0 and Isaac Sim v6.2.0 as current. Confirm its data units too; the state log lists joint data in radians [56]. The full metric list and the measured SO-101 data are in Sim-to-real metrics.

Reference figures

These are one paper's results, on its tasks, not targets [54].

Reference figures
FigureValueConditionSource
Real success, Squint on SO-10191.3 percentAveraged over 80 trials for 8 tasks[53]
Simulation success, Squint96.1 percentOver all tasks, at the end of its training run[54]

Gate D: classify, then fix

Log each gap with one class [8]. The class names come from the guideline. The placement of other guidelines under them is this book's.

Gate D: classify, then fix
ClassTypical checkSource
SensingNoise against the simulated model; camera assignment[57], [58]
ActuationFriction, backlash, thermal, delay[48], [25]
PhysicsContact behaviour against video[43]
ModelingCAD against as-built, mass against scale[51], [52]

Gate E: hardware test plan

Gate E: hardware test plan
LevelRequirementCheckSource
mustTest each policy through a systematic plan. Good simulation results do not show that it works on hardwaremeasure: run a written test plan[9]
shouldCheck input, system response and output consistency separatelyreview: report covers all three[59]
mustApply the same policy-output post-processing during real-world inference as in Isaac Labreview: diff both code paths[60]
shouldWhen a policy fails on the real robot, match the initial observations to simulation and compare responses. This separates observation mismatch from physics or controller mismatchmeasure: same observation sequence on both[61]
shouldClassify each gap as sensing, actuation, physics or modelingreview: gap log[8]
shouldPrefer a simple real controller interface. It has fewer parameters that can mismatchreview: compared against simpler options[62]

On the SO-101

On the SO-101
LevelRequirementDoc as the guideline states itSource
guideFirst transfer: consider scaling actions down and raising control rate. Squint used 0.15 and 30 HzSquint, arXiv:2602.21203; SO-101 in ManiSkill3[63]
shouldRecord any deliberate change between training and hardware rateSquint[64]
guide10 Hz delta joint position is Squint's training setting, chosen for faster trainingSquint[65]
shouldReport real success with its trial countSquint[53]
guideCite a simulation figure only with its conditionSquint[54]
mustCalibrate each arm before running a policyIsaac Lab 3.0.0 docs, SO-101 typing project (patched 3.0.0-based runtime)[66]
mustTyping: keep the keyboard fixed. The policy uses a fixed A–Z key map without camera input and cannot adjust to movementSame[27]
mustTyping: confirm on hardware that the arm presses and lifts off the right keysSame[67]
mustUse the typing project's pinned runtime. Unmodified 3.0 GA or later is not validatedSame[68]
shouldMoving a fixture or device invalidates the layout checkNVIDIA SO-101 sim-to-real course, latest alias, no version shown, captured 2026-11-08[28]
mustVerify camera assignments after any replug. Indices can changeSame course[58]
shouldKeep layout, lighting and props consistent across calibration, teleoperation and evaluationSame course[69]

Servo values to bound the fit, for the STS3215-C018 at 12 V: the database holds the maker-wiki no-load speed and backlash limit, and bench values for no-load speed, backlash and repeatability. This site does not show them. Take the bounds from the sheet of the servo you buy, or measure them on your arm. The follower uses 1/345 gearing [70].

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. Match action term order, joint names, width, target type, scale, offset and clipping exactly between backends. Level: must. isaac-sim.github.io 2026-11-08“Term order, ordered joint or body names, width, target type, scale, offset, and clipping.”
  2. Match observation groups, term order, width, history, units, frames, clipping and corruption exactly between backends. Level: must. isaac-sim.github.io 2026-11-08“Group and term order, tensor width, history length, units, frames, clipping, and corruption behavior.”
  3. Match physics dt, decimation, policy period and action hold; only Newton substeps may differ inside one policy period. Level: must. isaac-sim.github.io 2026-11-08“Physics dt, decimation, policy period, and action-hold behavior.”
  4. Confirm that the same task trains in both engines before attempting transfer. Level: should. isaac-sim.github.io 2026-11-08“ensure that the same task can be trained successfully in both engines.”
  5. Treat successful sim-to-sim transfer as a possible first step toward sim-to-real deployment. Level: guide. isaac-sim.github.io 2026-11-08“Successful transfer can be a first step toward sim-to-real deployment.”
  6. Fit armature, damping, friction, gains, inertia and delay to hardware logs with optimization, rather than hand-tuning, when matching a specific robot in MuJoCo. Level: should. mujoco.readthedocs.io 2026-11-08
  7. Report the gap per joint in position, velocity and torque, not as one robot-wide number. Level: should. docs.nvidia.com 2026-11-08
  8. Classify each observed gap as sensing, actuation, physics or modeling before choosing a fix. Level: should. docs.nvidia.com 2026-11-08
  9. Test each policy on hardware through a systematic plan; good simulation results do not show that it works there. Level: must. docs.nvidia.com 2026-11-08
  10. Raise damping when a position policy alternates saturated commands to exploit one solver's response. Level: should. isaac-sim.github.io 2026-11-08“Increased damping is often necessary to prevent bang-bang control.”
  11. Carry normalization statistics, recurrent state shape and reset, and commands unchanged with the checkpoint. Level: must. isaac-sim.github.io 2026-11-08“Normalization statistics, recurrent-state shape and reset, commands, and privileged inputs”
  12. Keep the mechanism the same between backends: body and joint lists in the same order, the same active joints and the same mimic or equality links. Level: must. isaac-sim.github.io 2026-11-08“Ordered bodies and joints, active degrees of freedom, and mimic or equality coupling.”
  13. Keep the episode definition the same between backends: how resets and commands are sampled, what rewards and terminations mean, episode length and the success test. Level: must. isaac-sim.github.io 2026-11-08“Reset and command distributions, reward and termination meanings, horizon, and success”
  14. Do not let a physics preset silently change policy-facing terms; if it does, treat the result as a different task. Level: must. isaac-sim.github.io 2026-11-08“A physics preset should not silently change policy-facing action, observation, reward, termination, command, or reset terms.”
  15. On branched robots, pick a single joint and body ordering convention and use it for the training run and for all playback, on either backend. Level: must. isaac-sim.github.io 2026-11-08“Set both joint_ordering and body_ordering to one convention for training and for every playback”
  16. Before any training or policy transfer, step one identical fixed state in each backend over several resets and compare the results. Level: should. isaac-sim.github.io 2026-11-08“Run the same fixed task state through both backends before training or transfer:”
  17. Expect similar behaviour, not identical trajectories, when moving a policy between physics backends. Level: should. isaac-sim.github.io 2026-11-08“Expect similar behavior, not identical trajectories.”
  18. Reproduce a Newton-only failure at its first failing step with one environment, fixed seed and reset state, no randomization, and the same actions in both backends. Level: should. isaac-sim.github.io 2026-11-08“Reproduce the first failing step with one environment, a fixed seed and reset state, no domain randomization, and the same action sequence in both backends.”
  19. During transfer, log how often actuators saturate and how often action signs flip between steps. Level: should. isaac-sim.github.io 2026-11-08“monitor saturation and consecutive action sign changes.”
  20. Do not port solver settings by number, because a configuration value has meaning only within its solver. Level: should. isaac-sim.github.io 2026-11-08“A configuration value has meaning only within the solver that consumes it.”
  21. If a transferred locomotion policy falls within a few dozen steps in every environment, rule out joint order before you blame contacts. Level: should. isaac-sim.github.io 2026-11-08“Rule ordering out before attributing such a collapse to contacts, friction, or actuator response.”
  22. On PhysX, check for doubled gripper effort when both mimic fingers have nonzero drives and receive one logical command. Level: must. isaac-sim.github.io 2026-11-08“effectively applying the command twice relative to MJWarp’s single driven finger.”
  23. Do not assume PhysX friction randomization carries over: Newton events use one friction coefficient, and PhysX values and buckets do not map one-to-one. Level: must. isaac-sim.github.io 2026-11-08“PhysX static/dynamic values and buckets do not map one-to-one.”
  24. Do not copy PhysX friction numbers into MJWarp; verify shapes and contact first, then fit friction to measured slip. Level: must. isaac-sim.github.io 2026-11-08“Do not copy PhysX friction settings numerically.”
  25. Suspect unmodeled delay when a policy shakes violently on hardware. Level: should. arxiv.org 2026-11-08
  26. Model joint friction where stiction exists, or the policy learns that commanded small motions are always reached. Level: should. isaac-sim.github.io 2026-11-08“the policy learns to expect that commanded joint positions are always reached.”
  27. Keep the SO-101 typing keyboard fixed: its policy uses a fixed A–Z key map without camera input and cannot adjust to movement. Level: must. isaac-sim.github.io 2026-11-08“it cannot adjust to a moved keyboard.”
  28. Moving any fixture or device invalidates the layout check, so verify positions again before the next run. Level: should. docs.nvidia.com 2026-11-08
  29. Expect mujoco.sysid to use box-constrained least squares, so give each parameter real bounds. Level: must. mujoco.readthedocs.io 2026-11-08
  30. Use the physically consistent pseudo-inertia helper when you identify link inertia with mujoco.sysid. 10-D pseudo-inertia Level: should. mujoco.readthedocs.io 2026-11-08
  31. Include sensor and actuator delays and biases in the identification when they exist. Level: should. mujoco.readthedocs.io 2026-11-08
  32. Budget one full rollout per parameter perturbation when planning a mujoco.sysid fit. Level: guide. raw.githubusercontent.com 2026-11-08
  33. To set stiffness and damping independently in MuJoCo, use the direct negative solref format. Level: should. mujoco.readthedocs.io 2026-11-08
  34. Combine identification with randomization; Hwangbo notes that accurate modeling is cumbersome alone and robustness alone can cost performance. Level: should. arxiv.org 2026-11-08
  35. Retune damping after increasing armature, because the effective natural frequency and damping ratio change. Level: should. isaac-sim.github.io 2026-11-08“Retune damping after increasing armature because the effective natural frequency and damping ratio change.”
  36. Measure control loop timing on hardware, because it differs from simulation. Level: should. docs.nvidia.com 2026-11-08
  37. Keep MuJoCo delay at or below timestep times nsample, because beyond that it extrapolates non-causally with no runtime error. Level: must. mujoco.readthedocs.io 2026-11-08
  38. Remember that the smallest effective positive MuJoCo delay is one timestep. Level: should. mujoco.readthedocs.io 2026-11-08
  39. For stochastic delay or identification, set nsample from the largest delay, and adjust actuator_delay at run time within it. Level: should. mujoco.readthedocs.io 2026-11-08
  40. For continuous-valued quantities in delay identification, use linear or cubic interp to avoid zero-order-hold discontinuity; not for integer-valued sensors. Level: should. mujoco.readthedocs.io 2026-11-08
  41. Know that mjlab replays actuation delay quantized to the physics timestep, so a fine real delay needs a fine timestep. Level: guide. arxiv.org 2026-11-08
  42. Plot a real step response to quantify stiction, then add joint friction in simulation; the UR10e showed about 0.25 degrees of offset. 0.25 degrees Level: should. isaac-sim.github.io 2026-11-08“observed contact offsets of approximately 0.25 degrees from the commanded setpoint.”
  43. Compare sim and real video of contact behaviour while tuning friction and solver values, without waiting for full training. Level: guide. isaac-sim.github.io 2026-11-08“Compare the gear’s behavior in the gripper visually between sim and real”
  44. Do not assume an ideal actuator model transfers on series elastic drives: on ANYmal, both ideal and analytical models failed. Level: must. arxiv.org 2026-11-08
  45. Consider a learned actuator model when hand-tuned analytical models stall; the ANYmal team spent over a week without much success. Level: guide. arxiv.org 2026-11-08
  46. Train an actuator network to predict joint torque from position error and velocity history, when analytical modeling is not good enough. Level: guide. arxiv.org 2026-11-08
  47. Check the independence assumption before one per-actuator network, because shared hydraulic supply may couple actuators. Level: should. arxiv.org 2026-11-08
  48. Model friction, backlash and thermal effects where they matter, because typical motor models lack them. Level: should. docs.nvidia.com 2026-11-08
  49. To measure the actuation gap, replay the same motions on hardware and in simulation and compare the two logs. Level: should. docs.nvidia.com 2026-11-08
  50. Use RMSE, MAPE, correlation and cosine similarity per joint when comparing with SAGE output. Level: guide. raw.githubusercontent.com 2026-11-08
  51. Compare CAD against the as-built robot as part of gap measurement. Level: should. docs.nvidia.com 2026-11-08
  52. This guideline is held back: its source does not allow us to publish it yet.
  53. Report real success with its trial count, as Squint does with 91.3 percent over 80 trials on 8 tasks, and show the simulation figure beside it. Level: should. arxiv.org 2026-11-08
  54. Cite a simulation success figure only with its training-time condition: Squint's 96.1 percent is measured at the end of its training run. Level: guide. arxiv.org 2026-11-08
  55. Check simulator versions before using SAGE, because its README badges and requirements name Isaac Sim 5.0.0, not 6.2.0. Level: must. raw.githubusercontent.com 2026-11-08
  56. Check units in SAGE data files, because the state log format lists joint data in radians. Level: should. raw.githubusercontent.com 2026-11-08
  57. Model sensor noise and delay in the observations where they matter for the robot. Level: should. isaac-sim.github.io 2026-11-08“Sensor noise and delays should be modeled appropriately”
  58. Verify camera assignments after any replug, because indices can change. Level: must. docs.nvidia.com 2026-11-08
  59. Check input, system response and output consistency separately before you call a transfer sound. Level: should. isaac-sim.github.io 2026-11-08“Successful sim-to-real transfer requires addressing three fundamental aspects:”
  60. Apply the same policy-output post-processing during real-world inference as in Isaac Lab. Level: must. isaac-sim.github.io 2026-11-08“is also applied during real-world inference”
  61. When a policy fails on the real robot, match the initial observations to simulation and compare responses to separate observation mismatch from physics or controller mismatch. Level: should. isaac-sim.github.io 2026-11-08“set up the real robot with the same initial observations as in simulation”
  62. Prefer a simple real controller interface, because it has fewer parameters that can mismatch between sim and real. Level: should. isaac-sim.github.io 2026-11-08“Have fewer parameters that can mismatch between sim and real”
  63. 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
  64. Record any deliberate change between training and hardware control rate; Squint accepts such a mismatch and states it. Level: should. arxiv.org 2026-11-08
  65. 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
  66. Calibrate each SO-101 before running a policy; another setup's released offsets do not calibrate your robot. Level: must. isaac-sim.github.io 2026-11-08“they do not calibrate another robot.”
  67. Before deploying SO-101 typing, confirm on hardware that the arm presses and lifts off the right keys. Level: must. isaac-sim.github.io 2026-11-08“checks that the real robot reaches and releases the intended keys in the trained scene.”
  68. Use the SO-101 typing project's pinned Isaac Lab runtime, because unmodified 3.0 GA or later is not validated. Level: must. isaac-sim.github.io 2026-11-08“Compatibility with unmodified 3.0 GA or later versions has not been validated.”
  69. Keep the real workspace layout, lighting and props consistent across calibration, teleoperation and evaluation. Level: should. docs.nvidia.com 2026-11-08
  70. The SO-101 leader uses several gear ratios, while the follower uses 1/345 gearing. Level: guide. huggingface.co 2026-11-08
  71. The Playground v1.0.0 SO-101 armature values identified from real teleop data are 0.1088 shoulder pan, 0.1472 shoulder lift, 0.1224 elbow, 0.1133 wrist flex, 0.1146 wrist roll and 0.1985 gripper. Level: guide. raw.githubusercontent.com 2026-11-09“# real teleop data. ARMATURE = { "shoulder_pan": 0.1088, "shoulder_lift": 0.1472, "elbow_flex": 0.1224, "wrist_flex": 0.1133, "wrist_roll": 0.1146, "gripper": 0.1985,”
  72. 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”
  73. Fit 24 parameters for the SO-101: armature, friction loss, stiffness and damping on each of the 6 joints, from teleoperation logs by rollout tracking error. 24 parameters Level: should. raw.githubusercontent.com 2026-11-09“The optimizer fits a 24-parameter model (`armature`, `frictionloss`, `stiffness`, `damping` for all 6 joints) from teleoperation MCAP logs by minimizing rollout tracking error in MuJoCo physics.”
  74. One supervised trial on a second SO-101 typed NEWTON in 7.10 s; this is one trial, not a success rate. Level: guide. raw.githubusercontent.com 2026-11-09“jointly fitted scene completed **NEWTON** in 7.10 seconds with matching releases and model-FK clearance,”
  75. The recorded physical result states that one NEWTON trial is not a success rate benchmark or an independent reproduction. Level: guide. raw.githubusercontent.com 2026-11-09“One operator-supervised NEWTON policy trial on the development setup; not a success-rate benchmark or independent-user reproduction.”
  76. The SO-101 pick-cube policy was tested on 3 robot setups; a visual gap can remain with darker shadows, so light the scene to avoid them. Level: should. raw.githubusercontent.com 2026-11-09“We tested on 3 different robot setups under various lighting conditions, but there may still be visual domain gap with darker shadows.”