justintiensmith/groot_multi_gpu_v2

GR00T-N1.7 Full Fine-Tune — Spa-Bench Epoch 12

This is the GR00T-N1.7 checkpoint evaluated as the Full Fine-Tune condition in Spa-Bench, a real-robot benchmark of spatially grounded reasoning.

Model details

Field	Value
Model repository	justintiensmith/groot_multi_gpu_v2
Base model	nvidia/GR00T-N1.7-3B
Checkpoint	End of epoch 12; step 76,596
Robot	SO-101 single-arm manipulator
Inputs	Fixed middle RGB, wrist RGB, six absolute joint positions, text instruction
Outputs	Six absolute joint-position targets
Action horizon	16
Adaptation	Language, visual, multimodal/projector, VLLN, and diffusion-action modules updated
Optimizer	AdamW, learning rate 1e-5, weight decay 1e-5
Schedule	5% warm-up, then cosine decay
Hardware and batch	Four NVIDIA GH200 GPUs; 24 samples per device, global batch 96

The embedded train_config.json pins the two-camera full-length training data to justintiensmith/VLA_Reasoning_Training_Dataset_1200_2cam@b82cdc8. It contains 1,200 episodes, 612,733 frames, 321 instruction strings, and the middle and wrist views used by this policy.

The run records the base-model identifier but not its immutable source revision. The current public GR00T base revision at archival review is 2fc962b973bccdd5d8ce4f67cc63b264d6886495; it must not be assumed to be the unrecorded training revision.

The author-supplied original train_groot_n17_full_ft.sh launcher is archived with the thesis artifact. It records seed 42, four-GPU training, the immutable dataset revision, component-tuning flags, checkpoint cadence, and the historical environment paths used for this run.

Deployment processing and intervention

Non-gripper action dimensions used the same causal filter as the Frozen LLM variant: filtered = 0.25 × current + 0.75 × previous_filtered. Filter state was initialized from the measured robot state and reset for every rollout.

This condition also received a small upward initialization assist before the scored timer. A retrospective estimate found 10.09 mm mean end-effector separation from the nominal start, including 7.63 mm mean upward displacement; the maximum paired arm-joint difference was 6.40°. Nominal start deviations were 9.27 mm for this condition and 3.09 mm for Frozen LLM. These values include ordinary reset/calibration variation and the intervention is a limitation when interpreting results.

Physical evaluation

The checkpoint completed 25/120 familiar/in-distribution trials (20.8%), with 20 trials from each Spa-Bench task family. Evaluation was stopped before the OOD and diagnostic protocol, so this checkpoint must not be compared with the fully evaluated policies on the headline OOD benchmark.

• Rollouts: justintiensmith/Spa_Bench_Partial_GR00T-N1.7_Full_Fine-Tune
• Thesis artifact: justintiensmith/Imperial-Thesis

Intended use and limitations

This release supports reproduction and analysis of the Spa-Bench experiment. Its evaluation is partial and contains no OOD trials. The reported outcome applies only to this checkpoint and protocol; it is not a general assessment of GR00T-N1.7. The initialization assist is unique to this condition and prevents a clean parameter-freezing ablation.

Robot policies can move hardware unexpectedly. Use conservative motion limits, an accessible emergency stop, a clear workspace, and direct supervision. Do not deploy this checkpoint for unattended or safety-critical operation.

Citation

Please cite the completed Spa-Bench MSc report, the thesis artifact, and the GR00T-N1.7 work referenced in the report.