What if you could run 4,000 robots at once on your laptop? That's the idea behind mjbatch, a Python library for running thousands of MuJoCo simulations in parallel on the CPU. It uses a C++ thread…

Co-designing a throwing arm and its controller together: link lengths, gear ratios, motor commands, and release time.

Cart-pole swing-up with predictive-sampling MPC. Roll out 1,024 candidate control sequences, pick the best, apply its first action, and repeat.

A G1 humanoid doing a backflip by imitating a reference motion. This is optimized with receding-horizon iLQR in 40 seconds on this same laptop.

I'm excited to release mjbatch, a simple tool that lets you step thousands of MuJoCo simulations in parallel on your CPU, all from the comfort of Python! To show how far this primitive can take you,…

A Go1 quadruped trained to follow joystick commands in 1024 parallel environments. It solves the task in 1 minute on my 5 year old M1 MacBook!