Stretch ICRA Run E

Private research checkpoint for the Hello Robot Stretch cup-serving policy experiments.

  • Experiment arm: Run E
  • Base architecture: NVIDIA GR00T N1.6-3B
  • Training data: GT60 + gate-passed pseudo45; corr >= 0.90 and 16-frame-window NMAE <= 25% (105 episodes)
  • Final training step: 78,420
  • Cameras: head + gripper
  • Action horizon: 16

Window-gated policy arm. The archived experiment used unstamped arm q01 statistics; preserve the included statistics for reproduction.

Files

This repository contains policy inference weights and the matching processor/statistics/configuration files. DeepSpeed optimizer and RNG states are intentionally excluded.

Use the included statistics.json (or processor/statistics.json, depending on the archived checkpoint layout) and processor configuration with the weights. The exact training configuration is retained under experiment_cfg/.

Intended use

Research and reproducibility only. Validate robot limits, preprocessing, normalization, and emergency-stop behavior before physical deployment.

Downloads last month
6
Safetensors
Model size
3B params
Tensor type
BF16
·
Video Preview
loading