Humanoid robot drives and drifts a go-kart via Direct Perception Control

机器之心 · wechat · 2026-08-17

Symbiosis Robotics demoed a humanoid that climbs into a cramped go-kart, steers through corners while coordinating throttle and brake, and even drifts at the finish. The core is Direct Perception Control (DPC): it removes the intermediate motion representation and separate Whole-Body Tracker, instead computing joint and hand targets directly from vision, task language and body feedback, with a two-stage DriftDistill training (offline BC init + online rollout correction) to handle real-world drift.

The team also curated 15,010 hours of embodied data (6,781h human egocentric, 4,024h armed robots, 3,660h wheeled humanoids, 545h bipedal), all converted into unified G1-executable joint actions. The write-up contrasts DPC with layered systems like Figure Helix, Gemini Robotics 2, ω-0 and Ψ₀, and shows additional demos of mobile pick-and-place and whole-body manipulation in confined spaces.

Original post →

More from Embodied

Embodied channel →