// the find
google-research/ravens
Train robotic agents to learn pick and place with deep learning for vision-based manipulation in PyBullet. Transporter Nets, CoRL 2020.
Ravens is the reference implementation for Transporter Networks (CoRL 2020), a Google Research paper on vision-based robotic pick-and-place in PyBullet simulation. It gives you 10 tabletop manipulation tasks, scripted oracles for imitation learning, and the Transporter architecture itself. This is squarely for robotics ML researchers reproducing or building on that paper.
1. The Transporter architecture is genuinely clever — it factors pick and place as two separate spatial attention problems (where to pick, then where to place conditioned on what was picked), which is much more sample-efficient than end-to-end approaches that treat the full action as one output. 2. Scripted oracles for all 10 tasks mean you get expert demonstrations for free, which is not trivial to build correctly for manipulation. 3. Task variety is good — deformable objects (rope, cloth), multi-step sequencing (Towers of Hanoi, pyramid), and generalization to unseen objects (kitting, packing) are all represented. 4. Pre-trained checkpoints and generated datasets are available, so you can evaluate without burning GPU time from scratch.
1. Frozen in 2020-era dependencies: Python 3.7, TensorFlow, CUDA 10.1/cuDNN 7.6.5. Getting this to run on a modern machine with current drivers is a real pain — pip will fight you. 2. PyBullet-only, and PyBullet itself is essentially unmaintained. Anyone doing serious work now is on Isaac Sim or MuJoCo; these tasks have no Isaac/MuJoCo ports in this repo. 3. No real-world transfer tooling — the paper validates on hardware but the repo is purely simulation with no bridge to ROS or real robot interfaces. 4. Last meaningful commit was 2024 but it's mostly just keeping the lights on — the core research code hasn't evolved and there's no path to 6-DOF tasks beyond the one underdocumented `transporter_6dof` agent.