| --- |
| license: apache-2.0 |
| tags: |
| - robotics |
| - imitation-learning |
| - diffusion |
| - flow-matching |
| - droid |
| --- |
| |
| # Gen2Act C35 Diffused Gripper |
|
|
| This repository contains the ep9 slim checkpoint for a Gen2Act policy trained |
| on DROID-derived robot demonstrations. The policy predicts an 8-step camera- |
| frame delta-action chunk from a source demonstration video and current robot |
| observations. |
|
|
| ## Model |
|
|
| - Architecture: fused-query flow-matching DiT |
| - Visual backbone: DINOv2 ViT-B/14 |
| - Action: 9D delta pose plus a diffused binary gripper dimension |
| - Inputs: 8 source frames, three-frame front and wrist observation histories, |
| current front patch geometry, and camera-projection proprioception |
| - Checkpoint: `c35_diffuse_gripper_fulltrain_ep9_latest_slim.pt` |
|
|
| The checkpoint is slim: it contains model weights, config, epoch, and metrics, |
| but no optimizer state. Load it with the matching Gen2Act codebase and config. |
|
|
| ## Training Notes |
|
|
| The model was trained for 10 epochs; this uploaded checkpoint is ep9. Its |
| training-side validation pose MAE was 0.03150 and pose RMSE was 0.06031 in the |
| normalized action-evaluation logging pipeline. These are not closed-loop robot |
| success metrics. |
|
|
| ## Intended Use and Limitations |
|
|
| This is a research checkpoint, not a safety-certified robot controller. It |
| requires calibration-compatible DROID-style observations and should be tested |
| in simulation or with appropriate hardware safety controls before any physical |
| deployment. |
|
|
