GR00T cloud robotics: V1, V2, and V3

Execution environment

Resource measurements

Training

Loss indicates optimization progress, not task success. Checkpoint selection must use validation evidence; final acceptance uses held-out seeds.

Failure analysis and corrective work

Earlier candidates were retained as negative evidence. Their failures were used to change the data, control path, and acceptance process before the final model was evaluated.

Corrections implemented

• Demonstration jaw target moved from mechanical bound -0.174 to 0.0 radians
• Removed artificial waypoint dwell
• Retained full release and retreat in every demonstration
• Policy preparation now happens before timed control on socket owner thread
• Errors identify joint and action index
• Optional native GR00T RTC prefix conditioning added; actual inference validation pending
• Lossless image compression with total decompression budget

Independent checks completed

• 53 local tests and 53 GPU tests pass
• V1, V2, and eight-step V2 checkpoints fully verified locally
• Final credential-free evidence archive contains 332 entries and is preserved with a SHA-256 checksum
• Offline processor restoration verified from 4.3MB credential-free bundle
• Failure and success videos saved with physics replay checks
• V3 passed 100-trial held-out production acceptance with 92 successes and zero joint-limit corrections
• Two real workstation-to-GPU campaigns completed 5 of 5 task successes each through an encrypted SSH tunnel

Research and implementation sources

• https://github.com/NVIDIA/Isaac-GR00T/tree/51d4c89f72fda44cbf77285c6a8114b52676b8a1
• https://huggingface.co/blog/async-robot-inference
• https://arxiv.org/abs/2506.07339

V1, V2, and V3 candidate comparison

Closed-loop run

Evidence: v1-runs\20260929T145456Z-87138805\episodes.jsonl. Sweep completed: False.

Checkpoint identity: 293b8cac12ae896728fc2285d6d7825e314f60facb97553498ef18befe6cb259. Prefetch lead: 0.25 seconds.

Closed-loop run

Evidence: v2-runs\20260929T160522Z-190d441b\episodes.jsonl. Sweep completed: True.

Checkpoint identity: 10a84840ba5f28b4588bab2b2eb95ba5e4c6e9325554dd596b91844dabe79030. Prefetch lead: 0.23 seconds.

Closed-loop run

Evidence: v3-runs\20260929T172231Z-f7aee121\episodes.jsonl. Sweep completed: True.

Checkpoint identity: 9bb09bc4fe31f86c72eef8f7e4821a8a176f25454b2bad164e1f9c4d1dfe97e5. Prefetch lead: 0.23 seconds.

Closed-loop run

Evidence: v3-validation\20260929T171645Z-d9da7813\episodes.jsonl. Sweep completed: True.

Checkpoint identity: 9bb09bc4fe31f86c72eef8f7e4821a8a176f25454b2bad164e1f9c4d1dfe97e5. Prefetch lead: 0.23 seconds.

Closed-loop run

Evidence: wan\20260929T175311Z-16639156\episodes.jsonl. Sweep completed: True.

Checkpoint identity: 9bb09bc4fe31f86c72eef8f7e4821a8a176f25454b2bad164e1f9c4d1dfe97e5. Prefetch lead: 0.75 seconds.

Closed-loop run

Evidence: wan\20260929T175609Z-50ee183c\episodes.jsonl. Sweep completed: True.

Checkpoint identity: 9bb09bc4fe31f86c72eef8f7e4821a8a176f25454b2bad164e1f9c4d1dfe97e5. Prefetch lead: 1.15 seconds.

Production acceptance

v2-runs\20260929T160522Z-190d441b\acceptance.json : FAIL

{
  "min_episodes": 30,
  "min_success": 0.8,
  "max_fallback": 0.05,
  "max_deadline_fraction": 0.05,
  "min_hz": 28.5,
  "max_projection_fraction": 0.01
}

• {"K": 40, "added_delay_s": 0, "backend": "real", "strategy": "deadline"}: excessive joint-limit projection

v3-runs\20260929T172231Z-f7aee121\acceptance.json : PASS

{
  "min_episodes": 30,
  "min_success": 0.8,
  "max_fallback": 0.05,
  "max_deadline_fraction": 0.05,
  "min_hz": 28.5,
  "max_projection_fraction": 0.01
}

Visual evidence

Recorded motion and outcomes

Run notes and remaining limitations

• V1 proved that the learned policy could complete the simulated task, but its action bounds and fallback behavior were not acceptable for release.
• V2 raised held-out task success to 98 percent, yet five trials exceeded the one percent per-episode joint-limit projection gate. The release was correctly rejected.
• V3 traded some headline task success for a safer learned distribution. It passed every declared gate with 92 successes in 100 fresh trials and zero joint-limit corrections.
• The real workstation-to-GPU route completed 10 of 10 tasks across baseline and tuned runs. Its 0.81 to 0.95 second p95 RPC latency left too many fallback ticks, so that Internet path is task-functional but not timing-qualified.
• The production approval applies to the declared MuJoCo SO100 task and checkpoint. Physical robot safety and sim-to-real performance remain separate validation work.

Interpretation

• A single simulated task with narrow initial-position randomization does not demonstrate general robotics capability.
• Command jerk and measured joint jerk are distinct. The reported jerk is computed from measured positions within each episode using actual physics timestamps.
• A 30-trial empirical 80% success threshold does not imply a 95% lower confidence bound of 80%. Confidence intervals are reported separately.
• CPU, rendering, RPC, and scheduling overhead are included in closed-loop timing. Offline visualization is excluded from timed control.