Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ManyQuadrupeds: Learning a Single Locomotion Policy for Diverse Quadruped Robots

Milad Shafiee, Guillaume Bellegarda, Auke Ijspeert

TL;DR

This study shows that drawing inspiration from animal motor control allows for a single locomotion policy capable of controlling a diverse range of quadruped robots, and evaluates the sim-to-real transfer by testing the single policy on both the Unitree Go1 and A1 robots.

Abstract

Learning a locomotion policy for quadruped robots has traditionally been constrained to a specific robot morphology, mass, and size. The learning process must usually be repeated for every new robot, where hyperparameters and reward function weights must be re-tuned to maximize performance for each new system. Alternatively, attempting to train a single policy to accommodate different robot sizes, while maintaining the same degrees of freedom (DoF) and morphology, requires either complex learning frameworks, or mass, inertia, and dimension randomization, which leads to prolonged training periods. In our study, we show that drawing inspiration from animal motor control allows us to effectively train a single locomotion policy capable of controlling a diverse range of quadruped robots. The robot differences encompass: a variable number of DoFs, (i.e. 12 or 16 joints), three distinct morphologies, a broad mass range spanning from 2 kg to 200 kg, and nominal standing heights ranging from 18 cm to 100 cm. Our policy modulates a representation of the Central Pattern Generator (CPG) in the spinal cord, effectively coordinating both frequencies and amplitudes of the CPG to produce rhythmic output (Rhythm Generation), which is then mapped to a Pattern Formation (PF) layer. Across different robots, the only varying component is the PF layer, which adjusts the scaling parameters for the stride height and length. Subsequently, we evaluate the sim-to-real transfer by testing the single policy on both the Unitree Go1 and A1 robots. Remarkably, we observe robust performance, even when adding a 15 kg load, equivalent to 125% of the A1 robot's nominal mass.

ManyQuadrupeds: Learning a Single Locomotion Policy for Diverse Quadruped Robots

TL;DR

This study shows that drawing inspiration from animal motor control allows for a single locomotion policy capable of controlling a diverse range of quadruped robots, and evaluates the sim-to-real transfer by testing the single policy on both the Unitree Go1 and A1 robots.

Abstract

Learning a locomotion policy for quadruped robots has traditionally been constrained to a specific robot morphology, mass, and size. The learning process must usually be repeated for every new robot, where hyperparameters and reward function weights must be re-tuned to maximize performance for each new system. Alternatively, attempting to train a single policy to accommodate different robot sizes, while maintaining the same degrees of freedom (DoF) and morphology, requires either complex learning frameworks, or mass, inertia, and dimension randomization, which leads to prolonged training periods. In our study, we show that drawing inspiration from animal motor control allows us to effectively train a single locomotion policy capable of controlling a diverse range of quadruped robots. The robot differences encompass: a variable number of DoFs, (i.e. 12 or 16 joints), three distinct morphologies, a broad mass range spanning from 2 kg to 200 kg, and nominal standing heights ranging from 18 cm to 100 cm. Our policy modulates a representation of the Central Pattern Generator (CPG) in the spinal cord, effectively coordinating both frequencies and amplitudes of the CPG to produce rhythmic output (Rhythm Generation), which is then mapped to a Pattern Formation (PF) layer. Across different robots, the only varying component is the PF layer, which adjusts the scaling parameters for the stride height and length. Subsequently, we evaluate the sim-to-real transfer by testing the single policy on both the Unitree Go1 and A1 robots. Remarkably, we observe robust performance, even when adding a 15 kg load, equivalent to 125% of the A1 robot's nominal mass.
Paper Structure (13 sections, 1 equation, 4 figures, 2 tables)

This paper contains 13 sections, 1 equation, 4 figures, 2 tables.

Table of Contents

  1. Introduction and Related Work
  2. Central Pattern Generators
  3. Learning Locomotion
  4. Contribution
  5. Learning Framework
  6. Action Space
  7. Observation Space
  8. Reward Function
  9. Results
  10. Implementation Setting
  11. Simulation Results
  12. Experimental Results
  13. Conclusion

Figures (4)

  • Figure 1: Simulation and experiment snapshots of training and deploying a single policy to control 16 different robots. Videos: https://miladshafiee.github.io/ManyQuadrupeds/
  • Figure 2: Characteristics and parameters of 16 diverse quadruped robots. These robots exhibit variations in mass, ranging from 2 to 200 kg, nominal height from 18 to 100 cm, and come in three different morphologies, with two types of DoFs (either 12 or 16).
  • Figure 3: Simulation results for training a single policy for 16 different quadrupeds: Top: Base locomotion velocity. Middle: CPG Frequencies for Front Right limb. Bottom: CPG Amplitudes for Front Right limb. We use the FR limb since the robots exhibit a trot gait where the feet closely repeat the same pattern.
  • Figure 4: Left: Snapshots of Go1 trotting on uneven grass. Right: Snapshots of A1 carrying $10$ to $15$ kg. The robot starts trotting with a $10$ kg mass, and then an additional $5$ kg is added. The robot successfully walks despite never having encountered any such disturbances during training.