Cross-Embodiment Manipulation Research — Robot Zoo

The Problem

One of the biggest bottlenecks in robot learning: data collected on one robot doesn't transfer to another. Almost no one has released clean, structured manipulation data across multiple robot embodiments in the same environment. Without this data, cross-embodiment policy transfer research is stuck.

What I Built

Set up three fundamentally different robots in simulation — Franka Panda (fixed-base, 7-DOF), Stretch 3 (mobile base, telescoping prismatic arm), and Fetch (mobile base, 7-DOF arm with torso lift). Made each perform the same manipulation tasks: pick cube, place cube, reach, push, and open drawer. Generated 1,000 scripted demonstrations per working robot-task combination and saved everything in LeRobot format for downstream training.

The Hard Part

The data generation wasn't the challenge — making each robot actually work was. Franka uses a standard IK controller and Warp state machine. Stretch's arm extends in the negative Y direction and requires precise base positioning before the arm moves. Fetch needs a specific grasp configuration just to get the gripper at the right height. Each robot required a completely different control strategy — which is exactly what makes the dataset research-relevant.

Engineering

Built a BaseRobot abstraction so all three robots share a common interface. Wrote 12 unit tests and 12 Jupyter visualization notebooks. Everything runs headless for batch generation, rendered mode for debugging. Full weekly documentation across 8 weeks including failures — more useful for reproducibility than most academic releases.

Results — Robot Breakdown

The Research Staircase