File size: 2,611 Bytes
05dbafa 724b4f7 05dbafa ce5f4e1 05dbafa ce5f4e1 05dbafa ce5f4e1 05dbafa 12f1b88 ce5f4e1 05dbafa 724b4f7 05dbafa 29a8b0a | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 | ---
license: apache-2.0
library_name: lerobot
pipeline_tag: robotics
tags:
- robotics
- act
- imitation-learning
- reinforcement-learning
- grpo
- mujoco
- bimanual
- roboeval
- hybird-control
datasets:
- helen9975/roboeval_ee_delta_20hz_scaled
- Sa74ll/LocalModels-ACT-bimanual-rollouts
---
# LocalModels / ACT-bimanual
Entry **ACT-bimanual** by team **LocalModels** (Saleh Alsharideh, Independent) to the CoRL 2026 RoboEval Diagnostic
Bimanual Challenge. **Official score 0.9873**, 399 of 400 episodes successful.
Seven tasks run hand-designed RGB-D controllers (no learned weights); LiftPot runs an Action Chunking Transformer (ACT)
trained on the demonstrations and fine-tuned in simulation with GRPO. Code, report and explainer videos:
[github.com/Sa74ll/RoboevalChallenge](https://github.com/Sa74ll/RoboevalChallenge).
| Path | Content |
|---|---|
| `submission/` | the submitted bundle: `metadata.yaml`, `policy/`, `checkpoint/` (seven `hybrid.json` controller selectors and the LiftPot ACT weights) |
| `official_run/` | the official run (seeds 1000-1049): `results.json` and the 400 episode logs |
| `SHA1SUMS.txt` | SHA-1 of every file in `submission/` |
The submission archive was packaged without `checkpoint/`; its `checkpoint_url` points to `submission/checkpoint` here.
## Run
Needs [RoboEval](https://robo-eval.github.io/challenge.html) at tag `corl2026-challenge`.
```bash
hf download Sa74ll/LocalModels-ACT-bimanual --include "submission/*" --include SHA1SUMS.txt --local-dir ACT-bimanual
cd ACT-bimanual && sha1sum -c SHA1SUMS.txt
pip install -r submission/policy/requirements.txt
python -m roboeval.challenge.evaluate --submission submission --workers W --gpus 0
```
## Results (official run)
| Task | Policy | Successes | Score |
|---|---|---|---|
| CubeHandover | controller | 50 | 1.000 |
| LiftPot | ACT | 50 | 1.000 |
| LiftTray | controller | 50 | 1.000 |
| PackBox | controller | 49 | 0.954 |
| PickSingleBookFromTable | controller | 50 | 0.991 |
| RotateValve | controller | 50 | 1.000 |
| StackSingleBookShelf | controller | 50 | 1.000 |
| StackTwoBlocks | controller | 50 | 0.953 |
| **Suite** | | **399 / 400** | **0.9873** |
Training data and every use of simulator state: the report, Sec. 3. The files here equal the submitted ones except
three edits that do not change behaviour (no contact e-mail in `metadata.yaml`, local paths shortened in the LiftPot
configs, a license notice in `modeling_act.py`); see [GitHub](https://github.com/Sa74ll/RoboevalChallenge).
## License
Apache-2.0. `modeling_act.py` is modified from LeRobot (Apache-2.0). |