| # Submission form draft |
|
|
| ## Artifact name |
|
|
| `scv-think-tank-robosyn-preliminary-v1` |
|
|
| ## Experiment name |
|
|
| `Task-specific TurboVLA + DP Hard450 preliminary submission` |
|
|
| ## Short description |
|
|
| Task-specific policy ensemble for the ten RoboSynChallenge manipulation tasks. |
| Six tasks use TurboVLA checkpoints warm-started from pretrained vision and text |
| encoders, click_bell uses a Diffusion Policy fine-tuned with 450 additional |
| hard-position trajectories, and three tasks use the strongest released official |
| checkpoint as a fallback. |
| |
| ## Data source |
| |
| Official released RoboSynChallenge simulation datasets for all ten tasks. The |
| click_bell policy additionally uses 450 hard-position trajectories generated |
| with the official simulator, task definition, and released assets. No external |
| demonstration dataset is used. Pretrained model weights are used for TurboVLA. |
|
|
| ## Run and dependency notes |
|
|
| The GitHub repository contains `eval.sh`, `deploy_policy.yml`, and |
| `deploy_policy.py` for each submitted policy adapter. The task-to-checkpoint |
| mapping and pinned official fallback revisions are recorded in |
| `task_checkpoints.json` in the Hugging Face repository. TurboVLA uses RGB |
| observations from cam_high, cam_left_wrist, and cam_right_wrist with a 50-action |
| horizon and ten environment steps per evaluation call. click_bell uses |
| Diffusion Policy with `dp_step=8`. Full setup and per-task commands are in the |
| GitHub README. |
|
|
|
|
