edge-inference/cohab-lekiwi

CoHab LeKiwi: two-robot floor-sorting skill data

Training data for the LeKiwi track of CoHab: two LeKiwi mobile robots (SO-101 arms) share one apartment and sort toys and fruit from the floor into a toy bin and a fruit bin. A vision-language model (Cosmos-Reason2-8B) picks each robot's next step; one pi0.5 policy executes the steps from each robot's own cameras. This dataset is what that policy is trained on.

• Simulator: ManiSkill 3 / SAPIEN, ReplicaCAD apartments 0-4 (apartment 5 is held out for evaluation and is not in the data), two robots per apartment, six objects (Duplo bricks, racquetball, wooden block, airplane part, plastic fruit).
• 19,293 episodes, 1,429,551 frames at 15 fps, 28 instructions. One episode is one step of one robot: "go to the ball", "pick up the ball", "bring the ball to the toy bin", "put the ball in the toy bin".
• LeRobot format v3.0 (LeRobotDataset("edge-inference/cohab-lekiwi")), robot_type: lekiwi, same keys and units as LeRobot's real LeKiwi.

How it was collected

Actions are always the teacher's. The teacher has simulator access: it drives along a planned floor path, grasps with an RL pick policy followed by a scripted lift, and holds the arm still while carrying. Only successful steps are kept, plus:

• short recovery segments (backing off from a mild failed pose), and
• 5,948 extra pick demonstrations started from where the learned policy tends to stop (88% of attempted picks succeeded; failed ones are not kept).

Features

key	shape	content
observation.images.front	240x320x3 video	low front camera RGB
observation.images.wrist	320x240x3 video	wrist camera RGB
observation.state	11	6 arm joint positions (deg, gripper 0-100), base x/y/theta velocity from the wheel encoders, goal x/y in the robot frame (m)
action	9	6 arm joint position targets, base x/y/theta velocity command (LeRobot LeKiwi convention)

The goal is where the current step's target (object or bin) is believed to be, from localization plus a prior map and camera sightings; in the data it carries 10 cm of noise. Depth is not included.

Simulator shortcuts to know

The base is moved by virtual x/y/yaw joints behind a model of the three wheel motors (LeRobot's kinematics, speed and acceleration limits, slip), not by simulated omni-wheel rollers; the chassis passes through rugs. Lighting is baked into ReplicaCAD's textures.

Results

A pi0.5 policy trained on this data (edge-inference/cohab-lekiwi, model) with Cosmos-Reason2-8B planning sorts 2.92 of 6 objects per task [95% CI 2.83, 3.02] over 1,000 tasks (random layouts) in an apartment not used for training (2,920 of 6,000 objects, 48.7%).

Code and full history: github.com/edge-inference/cohab-lekiwi (docs/PLAN.md). License: BSD 3-Clause.