mehul

I hand-wrote a 500-LoC RL stack to make hacking on RL research much easier. Most RL stacks are either massive and unhackable, or duct-taped research scripts. I am open-sourcing Mithrl, a modular RLVR stack. Next items on my checklist: adding more complex environment examples, supporting multi-gpu + async RL, and QoL fixes. I might scrap external runtime dependencies (Huggingface PEFT + vLLM) and write purpose-built, simpler versions from scratch if I feel the need. If you want to experiment with RL and are looking to own sovereign tools, I’d love to get on call, understand your requirements and help integrate for free.

Not so complicated now that you see how it can be made, is it? Such a satisfying process too😊

I ported a Torch-based differentiable simulator to pure JAX to run realistic cloth simulations.

Robot grasping has been fake for a long time. Vision only. No sense of touch. UMI-FT and CoinFT finally fix that.

You can now tell a robot “build me a chair” and it figures out the parts, the structure, and the assembly. Text to 3D to physical build. This paper quietly moves robotics from instructions to creation.

Control that isn’t coded but inspired from learning human movement. DreamControl teaches humanoid robots how humans move before teaching them tasks. They don’t start with control code. They start with intuition.

“A humanoid you don’t babysit.” Learns tasks from video. Handles objects it’s never seen. Costs less than an engineering student’s laptop. Trains overnight. Not another “demo” model. It’s an automation employee.

Robot learning is starved of data not only because it is hard, but because it is outright BORING! RoboCade: “what if we made it feel like Clash Royale instead of manual labor?”