README.md
| 1 | --- |
| 2 | license: mit |
| 3 | library_name: stable-baselines3 |
| 4 | tags: |
| 5 | - reinforcement-learning |
| 6 | - robotics |
| 7 | - myosuite |
| 8 | - gymnasium |
| 9 | --- |
| 10 | |
| 11 | # Joint-Space Empowerment for Dexterous Coordination in Tendon-Driven Hands |
| 12 | |
| 13 | This repository contains the official trained policies for |
| 14 | |
| 15 | > **Joint-Space Empowerment for Dexterous Coordination in Tendon-Driven Hands** |
| 16 | > *James Heald, Vittorio Caggiano, Vikash Kumar, Maneesh Sahani* |
| 17 | > **ICML 2026 (Spotlight)** |
| 18 | > [Paper](https://openreview.net/pdf?id=qI2eHwfNfh) | [Code](https://github.com/gatsby-sahani/JoSE) | [Project Page](https://joint-space-empowerment.github.io/) |
| 19 | |
| 20 | --- |
| 21 | |
| 22 | ## 📂 Repository Structure |
| 23 | |
| 24 | The files are organized by task environment and training seed: |
| 25 | |
| 26 | ```text |
| 27 | joint-space-empowerment |
| 28 | ├── Adroit |
| 29 | │ ├── BaodingBalls |
| 30 | │ │ ├── CustomAdroitBaodingP1-v1.yaml # Environment config |
| 31 | │ │ ├── seed_0 |
| 32 | │ │ │ ├── rl_model_10000000_steps.zip # Trained model |
| 33 | │ │ │ └── rl_model_vecnormalize_10000000_steps.pkl # Observation normalization statistics |
| 34 | │ │ ├── seed_1 … seed_4 |
| 35 | │ ├── DieReorient |
| 36 | │ │ └── ... |
| 37 | │ ├── KeyTurn |
| 38 | │ │ └── ... |
| 39 | │ └── PenTwirl |
| 40 | │ └── ... |
| 41 | ├── MyoHand |
| 42 | │ ├── BaodingBalls |
| 43 | │ │ └── ... |
| 44 | │ ├── DieReorient |
| 45 | │ │ └── ... |
| 46 | │ ├── KeyTurn |
| 47 | │ │ └── ... |
| 48 | │ ├── PenTwirl |
| 49 | │ │ └── ... |
| 50 | │ ├── Reorient100 |
| 51 | │ │ ├── Pretraining_Play |
| 52 | │ │ │ ├── seed_0 |
| 53 | │ │ │ │ ├── rl_model_replay_buffer_1000000_steps.pkl # Replay buffer |
| 54 | │ │ │ │ └── rl_model_vecnormalize_1000000_steps.pkl |
| 55 | │ │ │ ├── seed_1 … seed_4 |
| 56 | │ │ └── Training |
| 57 | │ │ └── ... |
| 58 | │ └── Reorient8-sparse |
| 59 | │ └── ... |
| 60 | └── README.md |
| 61 | ``` |
| 62 | |
| 63 | ## Loading Example |
| 64 | |
| 65 | > **Prerequisites:** clone and install the [JoSE repository](https://github.com/gatsby-sahani/JoSE). |
| 66 | |
| 67 | The `play.py` script in the JoSE repository downloads a pretrained model and executes it in the corresponding environment, allowing agent behaviour to be visualized. The script is called from the repo root as follows: |
| 68 | |
| 69 | ```bash |
| 70 | uv run python play.py --hand <hand> --task <task> --seed <seed> |
| 71 | ``` |
| 72 | |
| 73 | | Argument | Choices | |
| 74 | |---|---| |
| 75 | | `--hand` | `Adroit`, `MyoHand` | |
| 76 | | `--task` | `BaodingBalls`, `DieReorient`, `KeyTurn`, `PenTwirl`, `Reorient8-sparse`, `Reorient100/Training` | |
| 77 | | `--seed` | `0`, `1`, `2`, `3`, `4`, `5` | |
| 78 | |
| 79 | Seeds vary by hand and task. See the HuggingFace model directory to see which seeds are available for each task-hand combination. |
| 80 | |