---
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).