Light Robotics' Three Releases Chart a Path to Scalable Physical AI: Parkour, Navigation, Resilience
量子位 · wechat · 2026-09-13
A deep dive into Liangyuan Robotics' three consecutive releases—LightParkour, LightNav-0, and LightREACT—arguing the next phase of embodied AI is scaling capabilities across environments, tasks, and robot bodies.
- LightParkour: from a single human motion seed, physics simulation and curriculum learning expand skills up to 75cm obstacles (83% of the 90cm robot's height); multi-expert distillation merges walking and complex-contact skills into one unified vision policy running at 50Hz on-board.
- LightNav-0: a Real2Sim2Real engine converts 2,000+ internet scenes into reusable simulations yielding 4,000+ hours of VLA training data; PointCoT spatial reasoning lifts success by 8.4pp and SPL by 5.7pp across 8 ablations; the model zero-shot transfers to humanoid, quadruped, wheeled, and flying robots. Models, code, and tech report released.
- LightREACT: resilient whole-body control that infers body state from interaction history—limping, hopping, or crawling when joints are damaged, without fault labels.
The piece contextualizes these against SkildAI's S1 (66% vs 9% baseline), GeneralistAI's GEN-1.5 (59%→83% with fine-tuning), and RoboTTT's finding that 8K context improves closed-loop performance by 62%, framing context length as a new scaling dimension for robot foundation models.
More from Embodied
- Walden Robotics CTO: general-purpose robot manipulation is having its 'GPT moment' — adnothing · 2026-10-30
- Gym-Anything Team Stopped at 25 Simulations Due to Resource Limits, Seeks Collaborators — wellecks · 2026-10-30
- Gym-Anything Adds Audit Agents That Try to Break Simulations With Legal Robot Actions — wellecks · 2026-10-30
- NVIDIA researcher: train robot policies with privileged sim info, improving Pi 0.5 — wellecks · 2026-10-30
- Leveraging Simulator Privileged Information: Three Methods to Improve Robot Models — wellecks · 2026-10-30
- CMU Auto-Builds 25 Robot Simulations with Agents; Frontier Model Solves 0 of 25 Test Tasks — wellecks · 2026-10-30