justintiensmith/groot_multi_gpu_v4

GR00T-N1.7 Frozen LLM — Spa-Bench Epoch 12

This is the GR00T-N1.7 checkpoint evaluated as the Frozen LLM condition in Spa-Bench. “Frozen LLM” means that the language-model parameters were held fixed; the policy continued to receive language instructions.

Model details

Field	Value
Model repository	justintiensmith/groot_multi_gpu_v4
Base model	nvidia/GR00T-N1.7-3B
Checkpoint	End of epoch 12; step 71,304
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	LLM frozen; visual, multimodal/projector, VLLN, and diffusion-action modules updated
Optimizer	AdamW, learning rate 1e-4, 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 motion-trimmed training data to justintiensmith/VLA_Reasoning_Training_Dataset_1200_Trimmed_Start_5_Frame@c1231a8. That exact snapshot is a two-camera projection of the repository's five-camera trimmed release and is preserved by the groot-frozen-training-snapshot tag. The tabular trajectory file and middle/wrist videos are byte-identical between the two snapshots; episode metadata differs because the unused camera keys were removed. Both contain 1,200 episodes, 570,386 frames, and 321 instruction strings.

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

Deployment processing

For both GR00T variants, non-gripper action dimensions used the causal filter filtered = 0.25 × current + 0.75 × previous_filtered. Filter state was initialized from the measured robot state and reset for every rollout. The gripper command was excluded.

Physical evaluation

Condition	Successes	Rate
Familiar/in-distribution spatial instructions	81/120	67.5%
All withheld/OOD spatial configurations	133/300	44.3%
Matched OOD subset	50/120	41.7%
Matched direct-manipulation controls	98/120	81.7%

The matched-control gap was 40.0 percentage points. These are physical rollout results, not simulation metrics.

• Rollouts: justintiensmith/Spa_Bench_Full_GR00T-N1.7_Frozen_LLM
• Thesis artifact: justintiensmith/Imperial-Thesis

Intended use and limitations

This release supports reproduction and analysis of the Spa-Bench experiment. Older project files may call it “vision-only”; that term is inaccurate because several non-language modules were updated. The motion-trimmed data and deployment smoothing are policy-specific choices, so this comparison is not a controlled architecture 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.