Liangyuan Xinchuang Releases Three Physical AI Technologies; LightNav-0 Zero-Shot Transfers Across Four Robot Bodies

The three releases address different technical areas but are presented as answers to the same question: how to make Physical AI capabilities scale. Liangyuan Xinchuang describes a three-stage paradigm—Scalable Pre-Training, Scalable Alignment and Scalable Deployment. The stages are intended to form a continuing learning loop, in which large-scale data builds base capabilities, reinforcement learning and other methods improve alignment, and real-world deployment generates high-quality data that feeds further model optimization.

LightParkour focuses on parkour-style whole-body motion. According to the report, ordinary walking relies mainly on the legs and the ground for support, but climbing, vaulting or crossing higher obstacles makes the environment part of the body's support. Traditional motion imitation records how a person moves but not fully why that motion works in a given environment, and direct motion retargeting can produce hands floating in the air, the body passing through obstacles or trajectories that exceed a robot's dynamic limits. Collecting new motions for every obstacle also raises data costs with each skill, terrain and contact mode.

LightParkour starts from a short motion seed recovered from real human motion and places the motion and corresponding obstacle into physical simulation. The simulation re-establishes the relationship among the motion, obstacle, contact forces and robot actuator constraints. After the robot completes the current motion, the system raises the obstacle difficulty and uses successful trajectories as references for the next round of training. From an initial motion aimed at a 45 cm obstacle, the system can gradually generate reference trajectories covering obstacles up to 75 cm, about 83 percent of the height of the roughly 90 cm tall Lightbot 0. Only the initial motion and obstacle need manual alignment, with later capability growth driven by physical simulation and curriculum learning.

The report says LightParkour then uses multi-expert distillation to combine walking and three complex contact skills into a unified policy. Skill switching is also trained rather than governed by manual rules, so the robot learns when to keep walking, when to enter a whole-body motion and when to return to ordinary locomotion. During deployment, LightParkour runs as a unified recurrent depth vision policy on Lightbot 0. It uses only a chest-mounted depth camera, proprioception and velocity commands, runs at 50 Hz on onboard computing, and does not require runtime reference trajectories, external motion capture or off-board state estimation.

LightNav-0 addresses navigation. Traditional navigation systems can rely on pre-built maps, localization and calibration for a specific camera and robot body, the report says, while a foundation-model-oriented navigation system must handle different environments, tasks, viewpoints and robot bodies. The company's Real2Sim2Real data engine converts more than 2,000 real scenes from internet sources into reusable simulation environments, producing more than 4,000 hours of vision, language and action post-training experience. The same scene can generate different goals, routes, viewpoints and tasks, and camera geometry can change during training to avoid adapting only to one fixed robot viewpoint.

The report states that within the current experimental range, expanding environment coverage improved generalization more reliably than simply adding more trajectories in the same environment. This suggests that data scaling for Physical AI may differ from that of large language models: robots need more experiences, and they need those experiences across more diverse environments. LightNav-0 also introduces Point CoT. The model first predicts target points and traversable points in image space, then generates subsequent action tokens. The reasoning process changes from directly generating actions from vision and language to vision-language understanding, spatial reasoning and action generation. In eight public ablation evaluations, Point CoT raised average task success rate by 8.4 percentage points and SPL by 5.7 percentage points.

LightNav-0 then uses an RVQ action encoder to compress continuous robot trajectories into three action tokens, allowing vision, language, spatial position and robot actions to enter a unified training framework. The report says validation covered 10 simulation settings, including instruction following, object-goal navigation and embodied visual tracking. In real-robot deployment, the same model was zero-shot transferred to humanoid, quadruped, wheeled and flying robots. The company also released the model, code and technical report.

Light REACT is presented as work on whole-body resilience control under external disturbance, falls and hardware damage, studying how a robot can adjust its movement based on its own interaction history. The report links this to scaled deployment: robots will encounter abnormal states such as collisions, falls, actuator failures, locked joints and environments that restrict their usual motion. If such events frequently interrupt tasks and require human recovery, operating pressure rises as deployments increase. Light REACT is described as starting from this problem.

Across the three efforts, the report presents a shift in evaluation criteria. The previous stage asked how many skills a robot could demonstrate; the next stage asks whether those capabilities can continue to scale. For Liangyuan Xinchuang, that means models that can encounter more environments, cover more tasks, transfer to more robot bodies and adapt after entering the physical world to states not seen during training.