Growing Humanoid Parkour Skills through Real2Sim2Real

LightParkour uses humanoid parkour as a testbed for contact-rich skill learning. It grounds short human-motion clips in physics, grows each seed across terrain variation, and distills locomotion and whole-body skills into one deployable depth policy.

01

One policy across locomotion and parkour. A single onboard policy drives both locomotion and contact-rich parkour on Lightbot 0.

Follow the story Read the paper Watch on YouTube

A motion without its obstacle is only half a demonstration.

Parkour makes the central difficulty of contact-rich behavior explicit: the environment becomes part of the robot’s support structure. Success depends on where contact lands, when it bears load, and how the motion adapts as geometry changes.

Motion-tracking policies reproduce expressive whole-body skills but remain tied to the demonstrated scene. Reward-driven locomotion adapts to terrain, yet tends toward leg-dominant behavior. Neither alone produces geometry-aware, load-bearing contact.

Motion capture and ordinary video preserve movement, but rarely the scene that made it possible. Direct retargeting can leave hands floating, limbs intersecting obstacles, or trajectories the robot cannot execute.

Collecting a new demonstration for every obstacle restores the correspondence, but scales linearly with skill, contact pattern, and terrain. LightParkour instead recovers one short seed per skill and lets simulation generate the variation.

Simulation rebuilds the interaction, then expands it across terrain.

Real2Sim2Real turns each seed into a deployable skill family: physics restores valid contact, successful rollouts expand the terrain range, and distillation produces one recurrent depth policy.

  1. Real Recover interaction intent

    Whole-body timing and intended contacts from the source clip.

  2. Sim Ground it in physics

    Obstacle geometry, contact forces, and the robot’s actuator limits—then expansion across a range of obstacles.

  3. Real Deploy from perception

    One depth policy acts without a runtime motion reference.

02

Speed vault. Human motion → physics-grounded reference → onboard execution.

03

Reverse vault. Human motion → physics-grounded reference → onboard execution.

Physics repairs what retargeting breaks.

We retarget the source motion and place the obstacle where it implies load-bearing contact. The initial pair is physically invalid: hands sink, knees clip, and human dynamics exceed the robot’s limits. A tracking policy repairs the interaction under full physics, anchoring intended contacts and restarting from random reference frames to expose critical phases. The maneuver is preserved; the trajectory becomes executable.

04

Iterative reference refinement. From an obstacle-misaligned seed to terrain-paired, dynamically feasible references.

Each successful rollout raises the obstacle.

Each successful rollout becomes the next reference, feasible by construction under full physics and actuator limits. We raise the motion and obstacle by 5–10 cm and train again. Iteration grows one 45 cm climb into a continuous family through 75 cm—83% of the 90 cm robot’s standing height.

45 cm Seed obstacle
45–75 cm Generated reference range
0.83H Relative obstacle height

Only the initial motion–obstacle pair is aligned manually; physics and curriculum drive the remaining expansion.

The family compresses into one policy.

A motion seed specifies interaction intent, not a deployable trajectory. Simulation builds a family of feasible contact strategies. Privileged teachers learn those families; multi-expert and recurrent depth distillation compress them into one onboard policy.

05

The complete LightParkour training system. Reference generation grows terrain-paired references; multi-expert distillation unifies the skills; reward-only transition training learns the handoff; recurrent depth distillation produces a single 50 Hz onboard policy.

Multi-expert distillation consolidates the experts’ behaviors into one policy. Reward-only transition training then teaches it when to leave locomotion, enter a whole-body skill, and return. At deployment, depth and the velocity command drive those decisions—without a skill label or hand-authored switch.

During a vault, the chest camera may lose sight of the obstacle just as contact becomes load-bearing. Height-scan reconstruction trains the recurrent state to carry recent geometry through the occlusion. Training also models camera noise, dropout, 27–33 Hz frame rates, and 30–60 ms latency.

The policy has to survive four tests.

The evaluations test four requirements of a deployable contact-rich skill: physical feasibility, adaptation to geometry, composition across contact modes, and execution from onboard perception. All quantitative results are from simulation; hardware videos show transfer without real-world fine-tuning.

Fine-tuning and recurrence extend the deployable policy’s operating range.

Climb-and-step success across four obstacle heights, with 500 randomized simulation trials per setting.

90.0% at 70 cm · 0.77H
  • Full LightParkour
  • No final task fine-tuning
  • No recurrent memory
  • Privileged teacher · height scan
Privileged teacher 99.9%
Full LightParkour 99.2%
No final task fine-tuning 88.4%
No recurrent memory 54.0%

60 cm

0.66H

Privileged teacher 99.2%
Full LightParkour 98.8%
No final task fine-tuning 89.8%
No recurrent memory 56.6%

65 cm

0.72H

Privileged teacher 99.2%
Full LightParkour 90.0%
No final task fine-tuning 76.4%
No recurrent memory 34.2%

70 cm

0.77H

Privileged teacher 98.6%
Full LightParkour 33.4%
No final task fine-tuning 17.0%
No recurrent memory 0.0%

75 cm

0.83H · frontier

A 45 cm motion seed produces terrain-paired references through 75 cm. At the 75 cm frontier the deployable policy drops to 33.4% while the privileged height-scan teacher holds at 98.6%, exposing the perception-and-control gap. H denotes the robot’s 90 cm standing height.

Learned transitions turn isolated skills into continuous behavior.

+65 pp
Isolated skills No transition training
33.0%
Unified policy With transition training
98.0%
End-to-end transition-and-skill success rate across 100 randomized trials. A trial counts as successful only when the policy executes the correct handoff and completes the resulting segment.

Transition training opens a route between the two behaviors.

Multiple skills in one network do not create transitions by themselves. With training, the hidden state leaves locomotion, passes through transition states, and enters parkour when terrain demands it; without training, trajectories remain in locomotion. The view below links each state to a reconstructed pose. Read temporal paths and local neighborhoods—not global cluster distances.

06

A learned route through policy memory. GRU hidden-state trajectories during transition rollouts, colored by locomotion, transition, and parkour. Lines preserve rollout order; each selected point is grounded in its corresponding reconstructed pose. This qualitative view complements the 100-trial transition-success ablation above.

On hardware, only the deployable policy remains.

Lightbot 0 is our custom 90 cm, 18.9 kg humanoid with 21 actuated joints. A single recurrent policy runs onboard at 50 Hz from a chest-mounted depth camera, proprioception, and a velocity command—with no runtime reference, skill label, external tracking, or offboard state estimation.

1 deployable policy
50 Hz onboard control
0 runtime references

The policy times the handoff itself.

On mixed courses, the policy decides when to remain in locomotion, enter a whole-body skill, and return—without a runtime skill label or hand-authored switch.

07

Skill-to-skill. A direct handoff between two whole-body skills.

08

Skill-to-locomotion. A direct return to commanded locomotion.

09

Stair ascent and descent. Lightbot 0 uses onboard depth and the same deployable policy as the whole-body skills.

The same policy handles terrain between obstacles. On sparse footholds, it places each foot from support geometry that may already have left the camera’s field of view, relying on recurrent memory.

Parkour is the testbed. Contact-rich autonomy is the larger goal.

As an early test beyond parkour, we apply the same pipeline to a separate loco-manipulation policy. Trained with a rigid box, it also carries a compliant ball unseen during training—an encouraging but preliminary result across shape, mass, and contact dynamics.

12

Trained on the box, carrying the ball. No retraining for the new object.

Parkour is where we started, not where the method ends. The aim is a robot that can observe a contact-rich interaction once, practice across geometry and dynamics in simulation, and carry the resulting skill into the real world.

LightParkour

Versatile Perceptive Whole-Body Locomotion via Multi-Skill Distillation

LightParkour grows sparse contact-rich motion seeds into terrain-paired reference distributions, unifies locomotion and whole-body skills, and learns transitions from reward. The recurrent depth policy transfers to Lightbot 0 without real-world fine-tuning.

BibTeX
@misc{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},
  url    = {https://light-loco-parkour.github.io/}
}