Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Pedipulate: Enabling Manipulation Skills using a Quadruped Robot's Leg

Philip Arm, Mayank Mittal, Hendrik Kolvenbach, Marco Hutter

TL;DR

This work explores pedipulation - using the legs of a legged robot for manipulation using a dedicated pedipulation controller that is robust to disturbances, has a large workspace through whole-body behaviors, and can reach far-away targets with gait emergence, enabling loco-pedipulation.

Abstract

Legged robots have the potential to become vital in maintenance, home support, and exploration scenarios. In order to interact with and manipulate their environments, most legged robots are equipped with a dedicated robot arm, which means additional mass and mechanical complexity compared to standard legged robots. In this work, we explore pedipulation - using the legs of a legged robot for manipulation. By training a reinforcement learning policy that tracks position targets for one foot, we enable a dedicated pedipulation controller that is robust to disturbances, has a large workspace through whole-body behaviors, and can reach far-away targets with gait emergence, enabling loco-pedipulation. By deploying our controller on a quadrupedal robot using teleoperation, we demonstrate various real-world tasks such as door opening, sample collection, and pushing obstacles. We demonstrate load carrying of more than 2.0 kg at the foot. Additionally, the controller is robust to interaction forces at the foot, disturbances at the base, and slippery contact surfaces. Videos of the experiments are available at https://sites.google.com/leggedrobotics.com/pedipulate.

Pedipulate: Enabling Manipulation Skills using a Quadruped Robot's Leg

TL;DR

This work explores pedipulation - using the legs of a legged robot for manipulation using a dedicated pedipulation controller that is robust to disturbances, has a large workspace through whole-body behaviors, and can reach far-away targets with gait emergence, enabling loco-pedipulation.

Abstract

Legged robots have the potential to become vital in maintenance, home support, and exploration scenarios. In order to interact with and manipulate their environments, most legged robots are equipped with a dedicated robot arm, which means additional mass and mechanical complexity compared to standard legged robots. In this work, we explore pedipulation - using the legs of a legged robot for manipulation. By training a reinforcement learning policy that tracks position targets for one foot, we enable a dedicated pedipulation controller that is robust to disturbances, has a large workspace through whole-body behaviors, and can reach far-away targets with gait emergence, enabling loco-pedipulation. By deploying our controller on a quadrupedal robot using teleoperation, we demonstrate various real-world tasks such as door opening, sample collection, and pushing obstacles. We demonstrate load carrying of more than 2.0 kg at the foot. Additionally, the controller is robust to interaction forces at the foot, disturbances at the base, and slippery contact surfaces. Videos of the experiments are available at https://sites.google.com/leggedrobotics.com/pedipulate.
Paper Structure (25 sections, 3 equations, 7 figures, 1 table)

This paper contains 25 sections, 3 equations, 7 figures, 1 table.

Table of Contents

  1. INTRODUCTION
  2. RELATED WORK
  3. Manipulation via Locomotion
  4. Non-prehensile Pedipulation
  5. Prehensile Pedipulation
  6. METHOD
  7. Observation and Action Space
  8. Rewards
  9. Task Reward
  10. Normalization Rewards
  11. Terrain
  12. Command Curriculum for Loco-Pedipulation
  13. Sim-to-Real Transfer, Training, and Deployment
  14. RESULTS AND DISCUSSION
  15. Real-World Manipulation Skills
  16. ...and 10 more sections

Figures (7)

  • Figure 1: Our foot target tracking controller enables a variety of real-world manipulation tasks such as opening doors (A) and fridges (B), object transport (C), pressing a button (D), pushing obstacles out of the way (E), and collecting rock samples (F).
  • Figure 2: Overview of our training and deployment setup. We first specify the commands in an inertial frame, namely the world frame in training and the control frame during deployment. To make the policy agnostic to the used inertial frame, we transform the commands to the base frame before adding them to the observations. The policy's actions are interpreted as deviations from the current joint position.
  • Figure 3: Simulation setup in Isaac Gym. The initial and final sampling space of the foot position commands are visualized in green and blue, respectively. The command curriculum enables combined pedipulation and tripod locomotion in a single policy. The figure shows one robot tracking a close target point (left) and a robot approaching a far-range target point using the tripod gait (right). The target points are visualized in yellow.
  • Figure 4: Our controller enables numerous real-world manipulation tasks: (A) The robot opens a push door. (B) The robot opens a fridge. (C) The robot lifts a backpack and transports it to a box using a tripod-hopping gait. (D) The large workspace of the controller allows pressing a button far above the robot's base. (E) The controller can be used to push obstacles out of the way. (F) With an additional gripper, the robot can collect rock samples.
  • Figure 5: For close-range targets, the policy does not require a stance adaptation (A) and the tracking error quickly converges to a steady state value (C). For far-range targets, a tripod gait emerges (B), and the tracking error oscillates in the slow-down maneuver before converging (D).
  • ...and 2 more figures