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.
