Humanoids

Deep Whole-Body Parkour Puts Perception Into G1 Motion Tracking

Shar Hendrix 4 min read

Humanoid control usually splits into two camps: perceptive locomotion that walks over terrain but sticks to pedal gaits, and general motion tracking that copies complex skills on flat floors. Deep Whole-Body Parkour from Tsinghua’s Project Instinct tries to fuse both — whole-body motion tracking with exteroceptive sensing so dynamic, multi-contact skills survive uneven ground.

Paper: arXiv:2601.07701. Project: project-instinct.github.io/deep-whole-body-parkour. Video: YouTube.

Vaults, dive-rolls, kneel climbs, and depth-based corrections on a humanoid. Source: Project Instinct / YouTube.

What they claim

The abstract describes a single policy trained to perform multiple distinct motions across varied terrestrial features, with depth sensing in the control loop. Demo sections on the project page highlight:

  • Kneel climb
  • Dive roll
  • Roll vault
  • Jump sit
  • Auto position-correction from depth sensing
  • Robustness under distractors

Authors argue this expands traversability beyond walking or running into multi-contact parkour-style behavior on unstructured terrain.

Deep Whole-Body Parkour montage of vaults, dive-rolls, and outdoor climbs
Official demo montage for Deep Whole-Body Parkour. Source: Project Instinct YouTube (youtu.be/3s0Al0kxUuo).

Open pieces

The project lists open code for a related shadowing task in InstinctLab, motion-reference data for a G1 29-DoF torso-base model, and policy checkpoints. The broader Project Instinct site also ships InstinctLab, Instinct_RL, and Instinct Onboard repositories for whole-body control research.

Sibling work from the same group includes Embrace Collisions (CoRL 2025) and Hiking in the Wild (perceptive parkour framework, arXiv:2601.07718).

A Human’s Take

Parkour demos are the honesty test for whole-body policies: if the contact schedule only works on a flat mat, you do not have terrain intelligence. Depth-in-the-loop tracking is the piece worth watching — flat-floor acrobatics without perception already exists; surviving bad ground is the product problem.

Sources