SO-101 vial-to-rack: half-speed demonstration continuation
This is a complete, loadable GR00T N1.7 SO-101 policy checkpoint for the task
"Pick up the vial and place it in the rack." It continues the
reported 94% simulation checkpoint
at pinned revision bbfef9eaecc7a7654f42aaa4ceb95245bb140a53. The upstream
architecture is NVIDIA GR00T N1.7 3B.
The saved model configuration and both processor pipelines must be loaded with the
weights. The original SO-101 joint order, units, and camera keys are preserved.
Continued training
- Dataset: maximilianofir/so101_newton_dr_lr_half_20261002; 100 accepted scripted Newton simulation episodes, 50 left rack and 50 right rack, two 640 x 480 RGB views at 30 FPS.
- Scripted arm motion speed scale: 0.5 relative to the standard collection setting; gripper motion is not scaled.
- 80 episodes used for continued training. The common held-out validation set contains 20 episodes from this dataset and 20 from the other speed dataset.
- 4,000 optimizer updates; batch 32; action training horizon 32; AdamW learning rate 1e-5, 200 warmup updates, cosine decay; original image augmentation and saved processor statistics.
- Final logged training loss: 0.005447; full shared validation loss: 0.026236. W&B training history.
train_config.json records the saved training configuration. model.safetensors
SHA256: 90301e8aac2eba1ed83b17cb7de8294b71b2f89d5fff6cacd5ba191ad254da29.
Simulation evaluation
On the same 100 non-DR Newton scenes (50 left, 50 right; seeds 76000–76099), with both cameras, four reset render frames and 16 executed actions per prediction:
- 30-second limit: 59/100 (35 left, 24 right) successes.
- 60-second limit: 62/100 (35 left, 27 right) successes.
The starting checkpoint scored 42/100 at 30 seconds and 44/100 at 60 seconds on this balanced evaluation. Its previously reported 94/100 used an older, largely left-rack evaluation and is not directly comparable. Policy action sampling was not seeded per episode; scene-matched outcome differences are exploratory. The shared validation split holds out episode IDs, but scene independence across the two speed datasets could not be verified from the collection manifests. These datasets precede the later Newton retreat-gated scripted success and external-camera randomization changes; fresh datasets under the current code should be treated as a separate collection cohort.
Detailed results and limitations.
License
This fine-tuned derivative is distributed under the NVIDIA License included
in LICENSE. The base checkpoint's original NOTICE.txt is retained without
modification. The license limits use to non-commercial research or evaluation,
except as stated in its terms.