openarm-cube-in-box-v1

100 successful pick-and-place episodes recorded in MuJoCo with an OpenArm bimanual robot (left arm active, right arm parked). Task: "place the cube in the box". A 4 cm orange cube spawns at a random reachable position on the left half of a table; the robot picks it up and places it inside a fixed blue open-top box (16x16 cm inner).

Collected by Kael (Qualia Studios agent) for ACT-style policy training on the Qualia platform.

Format

LeRobot v3.0 dataset, fps 25, 2214 frames total.

Key	Shape	Notes
observation.images.wrist_left	480x480x3 video	left wrist fisheye cam (acting arm)
observation.images.wrist_right	480x480x3 video	right wrist fisheye cam (parked arm, sees robot body, contextual)
observation.images.base	480x480x3 video	base mast fisheye cam, full workspace overview
observation.state	float32 (9,)	left arm joint positions, rad: joint1..joint7, finger_joint1, finger_joint2
action	float32 (8,)	ABSOLUTE actuator position targets (MuJoCo data.ctrl[0:8]): joint1..joint7 targets, rad, plus gripper target (finger ctrl range)
task	string	"place the cube in the box"

Cameras

All three cameras render at 120 deg fovy in MuJoCo (pinhole) and are post-warped to an equidistant fisheye projection with OpenCV remap (black outside the fisheye circle). Two wrist cameras are mounted under the wrists; the base camera sits on a mast above the torso looking down at the workspace.

Collection

• Policy: chained PPO (stable-baselines3), grasp leg + place leg, switched after 10 consecutive control steps of stable grasped lift. Both legs MlpPolicy 256x256 trained on a delta-action env.
• Control: 100 Hz position deltas (max 0.10 rad/step). Frames captured every 4 control steps, so fps = 25. The recorded action is the absolute joint target commanded at the end of each 4-step window, which makes it directly usable as a position-control target for ACT.
• Success filter: only episodes ending with the cube settled inside the box (env check: cube xy within inner walls, z inside the box, speed < 5 cm/s) are kept. 100 successes out of 144 attempts (69.4%) with random spawn seeds.
• Episode length: 19 to 34 frames (mean 22.1), about 0.9 s of robot time per episode. The chained policy is fast.

Scene

• Cube spawn: x in [0.25, 0.33], y in [0.15, 0.38], rejection-sampled inside a reachability annulus r in [0.20, 0.42] around the left shoulder.
• Box: fixed at (0.35, -0.13) with +-1 cm jitter.
• Episodes start from a scripted hover pose: gripper open, 12 cm above the cube.

Caveats for training

• The wrist_left camera gets fully occluded by the cube during the grasp moment (extreme close-up). The base camera keeps an unoccluded workspace view at all times. Gemini judge verdict: PASS (3.5/5 usability).
• wrist_right mostly sees the parked right arm and torso; it carries little task information and is included for camera-slot completeness.
• The right arm is never actuated; observation.state and action cover the LEFT arm only.
• All data is simulation (MuJoCo 3.x), not real robot data.