DiffSim2Real: Deploying Quadrupedal Locomotion Policies Purely Trained in Differentiable Simulation

J. Bagajo, C. Schwarke, V. Klemm, I. Georgiev, J.P. Sleiman, J. Tordesillas Torres

2024 Conference on Robot Learning Workshop - CoRL Workshop, Munich (Germany). 09 November 2024

Summary:

Differentiable simulators provide analytic gradients, enabling more sample-efficient learning algorithms and paving the way for data intensive learning tasks such as learning from images. In this work, we demonstrate that locomotion policies trained with analytic gradients from a differentiable simulator can be successfully transferred to the real world.
Typically, simulators that offer informative gradients lack the physical accuracy needed for sim-to-real transfer, and viceversa. A key factor in our success is a smooth contact model that combines informative gradients with physical accuracy, ensuring effective transfer of learned behaviors. To the best of our knowledge, this is the first time a real quadrupedal robot is able to locomote after training exclusively in a differentiable simulation.

Publication date: 2024-11-09.

Citation:
J. Bagajo, C. Schwarke, V. Klemm, I. Georgiev, J.P. Sleiman, J. Tordesillas Torres, DiffSim2Real: Deploying Quadrupedal Locomotion Policies Purely Trained in Differentiable Simulation, 2024 Conference on Robot Learning Workshop - CoRL Workshop, Munich (Germany). 09 November 2024.