Software
Software arrives in layers, so that every hardware release is usable on the day it ships: first a driver that moves the hand, then teleoperation and force display, then the learning stack.
Packages
All code and simulation models are released under the MIT licence.
- Minimal driver · R1Servo ID setup, zero calibration, joint read and write, first-motion demoRepository
- Robot description · R1URDF, MJCF and USD models with a load-and-move check scriptRepository
- Fingertip sensor tools · R2Acquisition, module-level calibration and wrist-frame transformRepository
- Teleoperation · R2Leader–follower control with live hand and force-vector displayRepository
- Learning code · R3Training, deployment and evaluation for the policies in the paperRepository
- Pretrained policies · R3Checkpoints for the task suiteDownload
Simulation models
One hand description, exported to the three formats most simulators read. The wrist frame and the three fingertip sensor frames are defined as in the paper.
| Format | Use it with | Contents |
|---|---|---|
| URDF | ROS, PyBullet, Isaac Gym, general tooling | Links, joints, limits, inertias, visual and simplified collision meshes |
| MJCF | MuJoCo | As above, plus position actuators and fingertip force sites |
| USD | Isaac Sim, Isaac Lab | Articulation with drives and collision approximations |
What the models are for. They describe the hand: kinematics, workspace, grasping, policy prototyping. They do not model the removal tasks. Hot-melt bonds, Blu-Tack and granular media are hard to simulate faithfully, so results on the task suite are reported on the real robot.
Learning setup in the paper
An Action Chunking Transformer fuses vision, joint state, fingertip pose and wrist-frame force, and predicts joint-displacement chunks. There is no state machine, contact threshold or hand-written force controller.
| Observation | Size |
|---|---|
| RGB, two external and one wrist camera | 3 × 480 × 640 |
| Joint positions and velocities | 24 |
| Fingertip poses, wrist frame | 3 × 7 = 21 |
| Fingertip forces, wrist frame | 3 × 3 = 9 |
| Action: joint displacements | 12 |
| Setting | Value |
|---|---|
| Control rate | 15 Hz |
| Chunk predicted / executed | 100 / 50 steps |
| Demonstrations per task | 20 |
| Training | 60,000 steps, batch 8 |
| Inference latency | 23.5 ± 4.1 ms (RTX 2060) |