Light-Loco-
ParkourVersatile Perceptive Whole-Body Locomotion
via Multi-Skill Distillation

Hongming Chen, Zhuoran Li, Hongxi Wang, Jiangpeng Hu, Ziliang Li, Peize Liu, QingRui Zhao,

Xuhao Liu, Liang Pan, Ximin Lyu, Yuntao Ma^†, Tingxiang Fan^†

Light Origins

^† Robotics team co-leads

↗ Paper coming soon▶ Video coming soon

abstract

Humanoid robots should be able to recruit their entire body when cluttered terrain demands more than ordinary walking. Existing systems either track expressive references without terrain generalization, or react online while leaving the arms, torso, and knees largely unused.

Light-Loco-Parkour (LLP) is an end-to-end perceptive whole-body locomotion system driven only by onboard depth and a velocity command. A single deployable policy decides when to walk, balance, climb, step down, or vault—with no skill label, hand-coded gate, runtime motion graph, or reference input.

real-world highlights

One policy. Onboard sensing and compute. Indoor and outdoor deployment.

1deployable
policy
0runtime
skill labels
6+whole-body
terrains

method

From sparse motion seeds to one terrain-aware policy.

Stage 01 · Skill acquisition

Grow one seed into terrain-conditioned whole-body skills.

1
Sparse seeds. Begin each contact-rich skill from a single human-motion example.
2
Dynamic refinement. Turn the seed into physically feasible robot motion.
3
Terrain pairing. Expand motion and obstacle geometry together over a continuum.
4
Skill experts. Train robust, geometry-aware teachers for each behavior.

Stage 02 · Distill & deploy

Let perception choose the skill—without a state machine.

1
Multiple teachers. Combine locomotion and whole-body skill experts.
2
Unified student. Distill every behavior into a single depth-conditioned policy.
3
Reward-driven transitions. Learn when to enter and leave each skill online.
4
Onboard control. Run directly from a chest-mounted depth camera and velocity command.

Light-Loco-Parkour skill distillation, transition, and depth-policy pipeline

whole-body skills

One controller. Many ways through.

The robot adapts contact timing, posture, and force to terrain geometry, then returns to ordinary locomotion after the obstacle.

Stairs — up & downStable ascent and descent across changing step heights.

Narrow bridgePrecise foot placement over a slim beam.

Stepping stonesSparse footholds under onboard perception.

Skill-SkillTransition directly between two whole-body parkour skills.

Skill-LocoExit a whole-body skill and return smoothly to locomotion.

Loco-SkillEnter a whole-body skill directly from ordinary locomotion.

paper

Light-Loco-Parkour

Versatile Perceptive Whole-Body Locomotion via Multi-Skill Distillation.

Preprint links will be added when released.

@article{chen2026lightlocoparkour,
  title   = {Light-Loco-Parkour: Versatile Perceptive
             Whole-Body Locomotion via Multi-Skill Distillation},
  author  = {Chen, Hongming and Li, Zhuoran and Wang, Hongxi
             and Hu, Jiangpeng and Li, Ziliang and Liu, Peize
             and Zhao, QingRui and Liu, Xuhao and Pan, Liang
             and Lyu, Ximin and Ma, Yuntao and Fan, Tingxiang},
  year    = {2026}
}