SkildAI's humanoid played 140 virtual World Cup years of self-play to learn soccer from just a goal reward
量子位 · wechat · 2026-09-25
Embodied AI unicorn SkildAI unveiled "Messinator," a humanoid robot that learned soccer through self-play: after roughly 140 years of simulated matches in NVIDIA IsaacSim, it dribbles past defenders, shields the ball, intercepts, and takes real shots.
Key points:
- Reward was goals only: no hand-crafted rewards for dribbling, shielding, or recovering from falls — all emergent behaviors, echoing AlphaGo's self-play recipe in the physical world.
- Opponent = past self: new policy versions keep beating saved older versions, so each gain auto-generates a harder curriculum.
- Powered by SkildBrain/S1: a cross-embodiment foundation model that learns tasks from demo videos via in-context learning; prior omni-bodied training spanned 100K simulated bodies and 1,000 simulated years, letting the model recover balance after broken legs or added loads.
Founded in 2023 out of CMU's robotics ecosystem by professors Deepak Pathak (curiosity-driven exploration) and Abhinav Gupta, with talent from Meta, Tesla, and NVIDIA.
More from Embodied
- Guava distills GPT-5.4 into a 4B robot agent with just 2,000 simulated trajectories — jbhuang0604 · 2026-11-09
- Dex-One2Many learns dexterous manipulation from one human video, 95%+ success across 4 hands — jbhuang0604 · 2026-11-09
- Fly-by-Code: coding agents pilot drones via active visual and physical feedback loops — jbhuang0604 · 2026-11-09
- μ0 world model learns robot physics from video only, beats π0 by 5 points on RoboCasa365 — jbhuang0604 · 2026-11-09
- NVIDIA unveils GR00T 2 open-world robot action model, hitting 95% success with 5 demos — CyberRobooo · 2026-11-09
- CoRL keynote to explore open-source data and hardware for robotics — micoolcho · 2026-11-09