Offline reinforcement learning datasets collected on real TriFinger robots and with a simulated version of the robot. Trajectories for two dexterous manipulation tasks were collected: Pushing a cube to a goal position on the ground and Lifting it to a goal position and orientation in the air. Several dataset types collected with policies of varying proficiency are available. Pose estimates for the cube are part of the observations. Versions of the datasets including images from three cameras are available as well.
Use case
Automation of an Assembly Workstation
| Example observations | 9 joint positions q_t (rad) 9 joint velocities v_t (rad/s) 9 measured torques tau_t (Nm) Object pose (x, y, z, quaternion q_x, q_y, q_z, q_w) Object linear/angular velocity Global goal g(x, y, z) |
| Example actions | 9 torque commands c_t (Nm) applied to motors (bounded ±0.36 Nm) |
| Example reward | Push task: -|p_obj – g|^2 (distance to goal) – 0.1 |c_t|^2 (energy penalty) + 5.0 if |p_obj – g|^2 < 0.02m (success bonus) Lift task: Reward based on vertical distance to target height Bonus for keeping the object lifted |