so101-sim-grasp-v8a

Synthetic grasping demonstrations for the SO-101 arm, generated in simulation (experiment v8 of the so101-lab project). Task: pick up a single named object ("pick up the red cube"). The policy input is the stereo images and the arm's joint angles only (model A: the policy must infer the viewpoint itself); so101-sim-grasp-v8b contains the same episodes with the camera description appended to the state.

Model trained on this dataset: inkoziev/smolvla-so101-sim-v8a.

Generation

• Simulator: PyBullet, official SO-101 model (so101_new_calib.urdf), physical grasps (gravity, friction, torque-limited position control).
• Scene: 50×49 cm plywood board on a desk, textures made from photos of the real setup, random background clutter, mostly white arm, light mostly from behind the robot, soft object shadows, camera noise model fitted to the real camera (white balance, contrast, blur, sensor noise, JPEG).
• Object: one cube (22–35 mm) or ball (24–40 mm diameter), 12 colours, 12–24 cm from the shoulder-pan axis within ±30°, placed only where both camera eyes see it.
• Cameras: a stereo pair modelled on the real rectified USB stereo camera (6.2 cm baseline), widely randomised per episode: azimuth −45…+30° around the robot's front (+ = robot's left), 0.45–0.85 m from the pan axis, 0.15–0.60 m high, 15–55° looking down, roll ±4°, horizontal field of view 42–68°. Viewpoints with azimuth > 30° or distance > 0.85 m are held out of this dataset for evaluation.
• Demonstrations: scripted inverse-kinematics expert (approach, descend, close, lift, hold), lift as soon as the gripper is blocked by the object, sideways perturbations of the executed motion (1–3 cm) in half of the episodes with clean recorded actions.
• Servo model: shoulder_lift and elbow_flex stop short when moving in one direction (dead band proportional to gravity load, measured on a 7.4 V SO-101), commands delayed by 0–2 steps; 15 % of episodes have ideal servos. The expert compensates, so the actions contain the compensation.
• 2,000 episodes were generated; only successful ones are included. Episodes start from a random low ready pose with the gripper in view.

Features (30 fps)

Key	Shape	Content
observation.images.left, observation.images.right	360×640×3 (video)	rectified stereo pair
observation.state	6	joint positions (see units)
action	6	joint targets, same units as the joint part of the state
sim.phase	1	0 approach, 1 descend, 2 close, 3 lift, 4 hold
sim.flags	5	jaw_contact, fixed_contact, lifted, grasped, gripper_table_contact
sim.tcp, sim.target_pos	3	gripper and target positions (m, simulator frame)
sim.action_executed	6	action actually sent (differs in perturbed episodes)

Units: five arm joints in degrees (URDF angles, 0 = middle of the range, as LeRobot's calibration in degrees mode), gripper 0 = closed … 100 = open. The sim.* fields are ground-truth labels; standard policies ignore them. eval_seeds.json lists the scene seeds of the held-out episodes used for closed-loop evaluation.

Limitations

Simulation only: one object, no distractors, servo behaviour fitted to a single real arm. Simulator, generator and evaluation code: github.com/Koziev/so101-lab.