Perceptive Pedipulation with Local Obstacle Avoidance

TL;DR

This work introduces a reinforcement learning-based approach to train a whole-body obstacle-aware policy that tracks foot position commands while simultaneously avoiding obstacles, and shows that it generalizes to unknown environments with different numbers and types of obstacles.

Abstract

Pedipulation leverages the feet of legged robots for mobile manipulation, eliminating the need for dedicated robotic arms. While previous works have showcased blind and task-specific pedipulation skills, they fail to account for static and dynamic obstacles in the environment. To address this limitation, we introduce a reinforcement learning-based approach to train a whole-body obstacle-aware policy that tracks foot position commands while simultaneously avoiding obstacles. Despite training the policy in only five different static scenarios in simulation, we show that it generalizes to unknown environments with different numbers and types of obstacles. We analyze the performance of our method through a set of simulation experiments and successfully deploy the learned policy on the ANYmal quadruped, demonstrating its capability to follow foot commands while navigating around static and dynamic obstacles. Videos of the experiments are available at sites.google.com/leggedrobotics.com/perceptive-pedipulation.

Figures (10)

• Figure 1: Our perceptive pedipulation controller avoids obstacles with the pedipulating foot of a quadruped robot (A). It can also react to and avoid dynamic obstacles (B). By toggling the contact switch, the robot can push obstacles out of the way on command (C).
• Figure 2: High-level overview of the training process. The command spaces are denoted in pink, and the spawn spaces are denoted in cyan. In Stage 1, we learn obstacle-free pedipulation following arm2024pedipulate. Stage 2 introduces obstacles arranged in five scenarios, where the policy learns obstacle avoidance. Stage 3 increases the noise on the perceptive observations (H) to enable dealing with the noise and artifacts in the mapping pipeline during deployment (I).
• Figure 3: We use a height scan of size $2.4m \times 1.6m$ with a resolution of $0.1m$. The grid is shifted to the front by $0.2m$ to cover the reach of the pedipulating foot.
• Figure 4: We use a detailed collision model for the robot and train on rough ground to encourage the policy to take higher steps. The obstacles are simple cuboids.
• Figure 5: The blind pedipulation policy arm2024pedipulate tries to reach the foot position command (green). The obstacle in the path of the foot leads to a collision (red) and eventually to a critical failure of the policy (A). Our perceptive pedipulation policy can reach around a corner while avoiding collisions (B).