PGTT: Phase-Guided Terrain Traversal for Perceptive Legged Locomotion

TL;DR

Phase-Guided Terrain Traversal (PGTT) addresses robust perceptive legged locomotion by combining a robot-centric heightmap, per-leg phase encoding, and phase-driven reward shaping to enforce gait structure without IK or fixed gait priors. The method uses an asymmetric actor-critic setup and cubic Hermite splines to generate terrain-adaptive swing trajectories, learned via PPO in GPU-accelerated MuJoCo over curriculum-generated stair terrains with domain randomization. PGTT demonstrates higher survival under disturbances and discrete obstacles in simulation, achieves faster convergence than strong baselines, and transfers to a real Unitree Go2 with a real-time LiDAR heightmap pipeline, with preliminary cross-platform results on ANYmal C. The work suggests that phase-guided reward shaping is a simple, general mechanism to achieve robust perceptive locomotion across platforms.

Abstract

State-of-the-art perceptive Reinforcement Learning controllers for legged robots either (i) impose oscillator or IK-based gait priors that constrain the action space, add bias to the policy optimization and reduce adaptability across robot morphologies, or (ii) operate "blind", which struggle to anticipate hind-leg terrain, and are brittle to noise. In this paper, we propose Phase-Guided Terrain Traversal (PGTT), a perception-aware deep-RL approach that overcomes these limitations by enforcing gait structure purely through reward shaping, thereby reducing inductive bias in policy learning compared to oscillator/IK-conditioned action priors. PGTT encodes per-leg phase as a cubic Hermite spline that adapts swing height to local heightmap statistics and adds a swing-phase contact penalty, while the policy acts directly in joint space supporting morphology-agnostic deployment. Trained in MuJoCo (MJX) on procedurally generated stair-like terrains with curriculum and domain randomization, PGTT achieves the highest success under push disturbances (median +7.5% vs. the next best method) and on discrete obstacles (+9%), with comparable velocity tracking, and converging to an effective policy roughly 2x faster than strong end-to-end baselines. We validate PGTT on a Unitree Go2 using a real-time LiDAR elevation-to-heightmap pipeline, and we report preliminary results on ANYmal-C obtained with the same hyperparameters. These findings indicate that terrain-adaptive, phase-guided reward shaping is a simple and general mechanism for robust perceptive locomotion across platforms.

Figures (9)

• Figure 1: Real‑world example of the Unitree Go2 robot climbing stair terrain.
• Figure 2: Simulation snapshots of PGTT. Left: Go2 on stairs with projected front‑foot trajectories (red). Middle: Go2 traversing discrete obstacles. Right: ANYmal C on stairs.
• Figure 3: PGTT's robot-centric heightmap.
• Figure 4: PGTT combines curriculum learning, a robot-centric heightmap, reward shaping through Hermite splines, asymmetric actor-critic learning, and low-level PD controllers for effective perceptive legged locomotion.
• Figure 5: Distances relative to the hip-joint frame and world frame. The black leg is the nominal stance, the dashed line a possible swing trajectory, and the red leg a random leg configuration.