Robo Cat vs Objects on a Table

A simulated robo-cat that learned to fight objects on a table.
A side project, August 2026

The purpose was for me to get familiar with robotics. The robo-cat uses neural networks to walk and push objects off the table in a physically simulated world. Very useful.

GitHub

Click the table to place an object. Drag the scene to rotate. Scroll to zoom. Drag the cat to reposition it.

pause/resume
reset
automatic object placement
toggle unnecessary sound
adjust animation speed

Notes

16 actuated joints run and trained in MuJoCo.
A walker network (20k parameters) turns a state and velocity command into locomotion. Pre-trained.
A decider network (13k parameters) controls the walker to clear things off the table.
The walker was pre-trained before the decider, with a mirror-symmetry loss and a penalty on erratic motion.
The entire demo is 6.3MB.

References

Work	What it gave
Todorov et al., MuJoCo, IROS 2012	the physics engine
Schulman et al., Proximal Policy Optimization, 2017	the training algorithm
Rudin et al., Learning to Walk in Minutes, CoRL 2021	the walker’s reward and observations
Iscen et al., Policies Modulating Trajectory Generators, CoRL 2018	correct a scripted gait, don’t invent one
Siekmann et al., Periodic Reward Composition, RSS 2021	the gait clock and the foot rewards
Fu et al., Minimizing Energy Consumption, CoRL 2021	mechanical power instead of squared torque
Abdolhosseini et al., On Learning Symmetric Locomotion, MIG 2019	the symmetry loss that stopped the limp
Hoeller et al., ANYmal Parkour, Science Robotics 2024	a slow decider over a fast walker
Lecomte et al., Gait Analysis in the Cat, Front. Neuroinform. 2021	the joint angles the body is shaped to

License

MIT, except the physics engine: the demo embeds the MuJoCo WebAssembly build, Apache 2.0, whose notice ships beside it.

Full disclosure: I would have not been able to realize this project without agentic assistance. I learned a lot about managing this task, but I take no credit for the coding details.