BRISA Reading Group 09 - RoboStriker: Hierarchical Decision-Making for Autonomous Humanoid Boxing

​To close November we look at RoboStriker: Hierarchical Decision-Making for Autonomous Humanoid Boxing (Yin et al., 2026) from Shanghai Jiao Tong University.

​Most humanoid research is about the robot and the world: walking, balancing, picking things up. Boxing adds an opponent who is actively trying to knock you over.

​Multi-agent RL is the textbook framework for this kind of strategic interaction. Applied directly to 29 motors, though, it struggles: there are far too many ways to move that aren’t boxing or aren’t even standing.

​RoboStriker splits the problem in three. First, a motion tracker learns professional boxing moves from human motion capture on a Unitree G1. Those skills are then compressed into a compact latent space, projected onto a hypersphere so that every point in it is a physically plausible motion. Finally, two agents learn tactics through self-play, choosing points in this skill space instead of raw joint commands.

​First learn to move, then learn to fight. The result outperforms the baselines in simulation and transfers to the real robot.

​The same pattern keeps showing up in robot learning, from Play2Perfect to here. Learn a vocabulary of good movements first, then plan in it. How much does that limit what a robot can discover?

​Skim the abstract if you have ten minutes, watch the videos if you have two. Everyone is welcome - whatever you study, wherever you study and however much robotics you have done so far.

​Come with questions, disagreements, or your best guess at when we’ll see the first robot boxing match.

​🎟️ There are only a limited number of spots, so we can keep the session personal and interactive. Please register early and only attend if you’ve received a ticket.

​📍 Hosted at the Merantix AI Campus on December 7th. Doors open 18:30, we start at 18:45.

​👋 New here? Join the community group - everything else gets announced there first.