PUMA: Perception-driven Unified Foothold Prior for Mobility Augmented Quadruped Parkour

TL;DR

PUMA tackles perception-guided parkour for quadrupeds by learning a unified, end-to-end policy that uses onboard depth to infer an egocentric polar foothold prior $\mathbf{f}_t = \{ d_t^{(L)}, d_t^{(R)}, \psi_t, \psi_{t+1} \}$ to guide velocity tracking without explicit foothold tracking. It employs an asymmetric actor-critic with PPO, a regression estimator for $\mathbf{f}_t$, and a Probability Annealing Selection (PAS) schedule that gradually shifts from ground-truth to predicted footholds during training, promoting stable convergence. A Multi-Critic framework balances task, foothold, and style rewards, enabling coordinated learning of locomotion and foothold guidance across varied terrains. Experiments in simulation and on a real quadruped show robust sim-to-real transfer and agile performance on walls and stepping-stone terrains, validating the approach and its potential for scalable perception-driven locomotion.

Abstract

Parkour tasks for quadrupeds have emerged as a promising benchmark for agile locomotion. While human athletes can effectively perceive environmental characteristics to select appropriate footholds for obstacle traversal, endowing legged robots with similar perceptual reasoning remains a significant challenge. Existing methods often rely on hierarchical controllers that follow pre-computed footholds, thereby constraining the robot's real-time adaptability and the exploratory potential of reinforcement learning. To overcome these challenges, we present PUMA, an end-to-end learning framework that integrates visual perception and foothold priors into a single-stage training process. This approach leverages terrain features to estimate egocentric polar foothold priors, composed of relative distance and heading, guiding the robot in active posture adaptation for parkour tasks. Extensive experiments conducted in simulation and real-world environments across various discrete complex terrains, demonstrate PUMA's exceptional agility and robustness in challenging scenarios.

Figures (6)

• Figure 2: Overview of PUMA training framework. A velocity-tracking locomotion policy takes proprioception and depth images as input to predict egocentric foothold priors, base velocity, and latent terrain features. These representations are then concatenated with the current observations and fed into the policy network. Multiple critic networks are trained on distinct reward components to cooperatively optimize the policy. During training, a PAS strategy gradually replaces ground-truth footholds with predicted ones. The entire process is conducted in a single stage, with all networks optimized simultaneously.
• Figure 3: Terrain difficulty gradually increasing from the left side towards the right. Notably, we introduce height variations to horizontal surfaces to simulate roughness, while the inclined walls remain smooth.
• Figure 4: Temporal evolution of body velocity and total contact force during a complete jump motion. The four motion phases are visualized in the top row and delimited by vertical green markers. The curves compare the body velocity of PUMA (orange) and the w/o MuC (blue) against the reference velocity (red dashed line), while the bottom bars represent the corresponding total contact forces.
• Figure 5: Real-world experimental results: (left) Wall-assisted Gap terrain, featuring stepping walls at 60° and 80° angles followed by a 1.2 m wide gap; (center) Stepping Stones terrain composed of discrete stones 0.5--0.8 m in length/width and 0--0.4 m in height variation; (right) Surmounting terrain with stepping walls at 60° and 80° angles leading to a 0.7 m high platform.
• Figure 6: Training curves analysis. (a) PUMA obtains significantly higher rewards than the w/o MuC baseline in foothold group. (b) Comparison of learning efficiency with different annealing schedules, where 5K and 8K denote that the annealing process concludes at the corresponding training steps.