Organizations: Noetix Robotics, Beijing, China. · the University of Science and Technology of China (USTC), Hefei, China. · [Institution Name], [City], [Country]. · the National University of Defense Technology (NUDT), Changsha, China.
We present a progressive imitation framework that learns terrain-appropriate humanoid parkour skills and subsequently adapts their hand--knee support. Human demonstrations provide traversal behaviors, but their contact patterns need not suit a robot's proportions or the available support surfaces. Our framework changes the role of demonstrations across three stages: privileged imitation produces physically executable robot references; terrain-indexed adversarial motion priors guide skill acquisition; and geometric support guidance with a reduced motion-prior weight enables contact adaptation. Candidate-based precontact attraction and actual-contact evaluation guide support without prescribing a contact sequence. The deployed policy selects behaviors from onboard depth, proprioception, and velocity commands without an external skill label. On the Noetix E1, the policy achieves 96.34% hybrid success in simulation and transfers without real-world policy fine-tuning, completing 29 of 30 individual obstacle trials and four of five hybrid courses. Support-adaptation ablations show higher traversal success for the complete adaptation stage on the tested terrains, while qualitative observations show changes in contact placement. These results support acquiring skills from demonstrations before relaxing their influence to accommodate robot-specific support configurations.
Figures & tables
Fig. 2: Overview of the training framework. Motion preprocessing (right) retargets human demonstrations and physically refines them into robot reference rollouts with kinematic, height-scan, and dynamic information. Terrain-indexed adversarial supervision compares reference and policy motion windows to provide a routed style reward. Geometric hand–knee candidates (bottom) guide precontact placement, while contact evaluation and state transitions determine support and retention terms. Task, style, and support rewards are combined with the weights in Eq. 1 for actor–critic training. The actor and critic use gated expert networks, distinct from the discriminator heads. At deployment, only the actor and its observation-processing modules are retained; privileged geometry, reference motions, discriminators, and the critic remain in training.
Fig. 3: Support Candidate mechanism (schematic, not to scale). (a) The frame’s open center is excluded from support surfaces. (b) Footprint compatibility and approximate 3-D ellipsoidal reachability filter candidate sets; the dashed ellipse illustrates a projection. Filled circles with rectangles denote valid candidate footprints, a cross denotes edge rejection, and an open circle denotes an unreachable candidate. (c) Predicted endpoints receive weighted attraction to valid candidates. (d) Actual contact quality is queried at the measured endpoint, with contact, candidate-availability and phase gates. The requested band width is 0.55 m; the native horizontal top is approximately 0.45 m wide.
Parameter
Value
Parallel environments / rollout
4,096 / 24 control steps
PPO epochs / minibatches
5 / 4
PPO learning rate / target KL
10−3 adaptive / 0.01
Clip / discount / GAE
0.2 / 0.99 / 0.95
Entropy / value loss / max. gradient
0.01 / 1 / 1
Discriminator learning rate
5×10−6
TABLE II: Training configuration shared by the formal comparisons.
Method
Low
Mid
High
Hybrid
OOD
Ours
100
98.44
98.14
96.34
93.46
w/o Stage 2
41.36
21.14
13.87
0
0
w/o Stage 3
100
85.55
79.69
83.59
85.11
Velocity Tracking
96.09
0
0
0
0
Single AMP
100
65.82
0
0
44.29
Multi-Skill Distillation
100
99.28
95.10
79.65
65.33
TABLE III: Complete-traversal success (%): 2,048 episodes per condition. Low/Mid/High follow Table I . Hybrid requires all three obstacles without resets; OOD heights lie 1–3 cm outside the corresponding training ranges.
Fig. 4: Supplementary evidence for terrain–skill correspondence. Top: a simulation montage of sequential obstacle traversal. Middle: joint-angle DTW distances (rad) to the High, Mid, and Low references. Bottom: lowest-distance intervals, with the dominant sequence High → Mid → Low. The interval labels identify reference matches rather than independently measured terrain boundaries.
Fig. 5: Qualitative support comparison. Top: w/o Stage 3; bottom: Ours. Each row combines a simulation contact close-up (left) and hardware traversal snapshots (right). Red circles in the upper row highlight hand-contact regions near the platform edge. The comparison illustrates support adaptation rather than a measured contact-force or stability margin.
Method
Low
Mid
High
1m×1m Top
Ours
100
98.44
98.14
93.46
w/o Candidates
99.17
83.15
71.53
46.24
w/o Relaxation
99.56
85.40
83.20
58.89
w/o Stage 3
100
85.55
79.69
50.63
TABLE IV: Success rates for support-adaptation ablations (%).
Fig. 6: Supplementary real-world hybrid traversal sequence. Snapshots are ordered from left to right, top to bottom, with timestamps relative to t0 . The two rows show different obstacle-crossing maneuvers within the same trial, illustrating terrain-dependent skill switching by a single perception-conditioned policy without external skill-selection commands.
Achieving highly dynamic parkour with humanoid robots on complex terrains unseen during the pre-training stage remains a challenge in robotics. Although general locomotion policies demonstrate capabilities across broad terrain distributions, they often struggle with arbitrary and highly challenging environments. To overcome this limitation, we propose a real-to-sim-to-real framework for rapid test-time training (TTT): environment-specific policy fine-tuning in simulation before deployment on a newly encountered terrain. We adopt a two-stage end-to-end learning paradigm: a policy is first pre-trained on diverse procedurally generated terrains, followed by rapid fine-tuning on high-fidelity meshes reconstructed from real-world captures. Specifically, we develop an efficient, feed-forward, and high-fidelity geometry reconstruction pipeline using RGB-D inputs, ensuring both speed and quality during test-time training. We demonstrate that TTT-Parkour empowers humanoid robots to master complex obstacles, including wedges, stakes, boxes, trapezoids, and narrow beams. The entire pipeline, including capture, reconstruction, and test-time training, takes approximately 10 minutes for most tested terrains. Experiments show improved traversal on the tested terrains after fine-tuning, with direct sim-to-real deployment and no further policy updates on the robot.
Shaoting Zhu, Baijun Ye, Jiaxuan Wang +5
Tsinghua University · Shanghai Qi Zhi Institute · Princeton University
While recent advances in humanoid locomotion have achieved stable walking on varied terrains, capturing the agility and adaptivity of highly dynamic human motions remains an open challenge. In particular, agile parkour in complex environments demands not only low-level robustness, but also human-like motion expressiveness, long-horizon skill composition, and perception-driven decision-making. In this paper, we present Perceptive Humanoid Parkour (PHP), a modular framework that enables humanoid robots to autonomously perform long-horizon, vision-based parkour across challenging obstacle courses. Our approach first leverages motion matching, formulated as nearest-neighbor search in a feature space, to compose retargeted atomic human skills into long-horizon kinematic trajectories. This framework enables the flexible composition and smooth transition of complex skill chains while preserving the elegance and fluidity of dynamic human motions. Next, we train motion-tracking reinforcement learning (RL) expert policies for these composed motions, and distill them into a single depth-based, multi-skill student policy, using a combination of DAgger and RL. Crucially, the combination of perception and skill composition enables autonomous, context-aware decision-making: using only onboard depth sensing and a discrete 2D velocity command, the robot selects and executes whether to step over, climb onto, vault or roll off obstacles of varying geometries and heights. We validate our framework with extensive real-world experiments on a Unitree G1 humanoid robot, demonstrating highly dynamic parkour skills such as climbing tall obstacles up to 1.25m (96% robot height), as well as long-horizon multi-obstacle traversal with closed-loop adaptation to real-time obstacle perturbations.
Training perceptive humanoid locomotion policies that traverse complex terrains with natural gaits remains an open challenge, typically demanding multi-stage training pipelines, adversarial objectives, or extensive real-world calibration. We present PRIOR, an efficient and reproducible framework built on Isaac Lab that achieves robust terrain traversal with human-like gaits through a simple yet effective design: (i) a parametric gait generator that supplies stable reference trajectories derived from motion capture without adversarial training, (ii) a GRU-based state estimator that infers terrain geometry directly from egocentric depth images via self-supervised heightmap reconstruction, and (iii) terrain-adaptive footstep rewards that guide foot placement toward traversable regions. Through systematic analysis of depth image resolution trade-offs, we identify configurations that maximize terrain fidelity under real-time constraints, substantially reducing perceptual overhead without degrading traversal performance. Comprehensive experiments across terrains of varying difficulty-including stairs, boxes, and gaps-demonstrate that each component yields complementary and essential performance gains, with the full framework achieving a 100% traversal success rate. We will open-source the complete PRIOR framework, including the training pipeline, parametric gait generator, and evaluation benchmarks, to serve as a reproducible foundation for humanoid locomotion research on Isaac Lab.
Chenxi Han, Shilu He, Yi Cheng +2
Tsinghua University · ZERITH Robotics · Shanghai University