gr00t-n1.7-ego-cloudwalk-100k

GR00T N1.7 — Ego + CloudWalk — 100k

Inference-ready checkpoint from a 100,000-step mixed-data fine-tune of NVIDIA GR00T N1.7.

Training mix

• 65% converted tabletop egocentric data
• 35% CloudWalk robot data
• CloudWalk split: 90% train, 10% held out
• Global batch size: 48 on 8 GPUs
• MXFP8 compute with FP8 gradient transport
• Final training loss: 0.0068698912

The objective was to retain performance on the bottle-pickup teleoperation task while introducing substantially more diverse egocentric tabletop behavior.

Files

This repository contains the consolidated inference checkpoint only. DeepSpeed optimizer states are intentionally excluded.

Evaluation

The checkpoint was evaluated with the frozen CloudWalk-v10 behavior probe: 48 action chunks from 16 episodes, prompt grab the bottle, seed 0, four denoising steps, and complete 40 x 64 SONIC outputs. Probe-manifest SHA-256: f05f95a9f4b399d8859ded9dc4a3504fcbd6be095c08d89d6be26a184f28c9d2.

These values measure SONIC-token similarity on the frozen bottle-task probe; they are not physical robot success rates. The 100k checkpoint retains essentially the same bottle-task behavior as the 65k checkpoint while training for another 35k mixed-data steps.

This is a research checkpoint and has not been validated for safety-critical robot deployment.