abhi-rf/gr00t-n1.7-ego-cloudwalk-10k

GR00T N1.7 ego + CloudWalk checkpoint (10k)

Private, inference-ready checkpoint from the mixed-data experiment at optimizer step 10,000.

Training mixture

• 65% retargeted tabletop ego data
• 35% CloudWalk bottle teleoperation data
• Global batch size 48 on 8 GPUs
• Selective MXFP8 action head with FP8 ZeRO-2 gradient transport
• AdamW, peak learning rate 1e-4, 5,000-step warmup, cosine decay

The training objective is to preserve bottle-pickup performance while adding diverse ego-derived motion data to reduce overfitting and catastrophic forgetting.

Frozen CloudWalk probe at 10k

The fixed protocol uses 48 action chunks from 16 CloudWalk-v10 episodes, prompt grab the bottle, seed 0, four denoising steps, and complete 40x64 SONIC tensors.

• Centered cosine versus CloudWalk-v10 30k: 0.9856
• Linear CKA versus CloudWalk-v10 30k: 0.9856
• Centered cosine versus recorded targets: 0.9822
• Centered cosine versus NVIDIA base + SONIC metadata adapter: 0.5814

These are open-loop, in-distribution token-similarity results; they do not by themselves establish closed-loop or physical pickup success.

Contents

This repository contains model shards, processor/configuration metadata, statistics, trainer state, and experiment configuration. DeepSpeed optimizer/rank state is intentionally excluded because it is not required for inference or downstream fine-tuning from the consolidated model weights.

Use is subject to the upstream NVIDIA GR00T model terms and applicable dataset terms.