DiffSim2Real: Deploying Quadrupedal Locomotion Policies Purely Trained in Differentiable Simulation

2024/11/04 by Joshua Bagajo, Bagajo, Joshua, Clemens Schwarke +13 · 3 citations
Computer Science · Engineering · #Evacuation and Crowd Dynamics #FOS: Computer and information sciences #Human Motion and Animation #I.2.9 #I.6.5 #Machine Learning (cs.LG) #Music Technology and Sound Studies #Robotics (cs.RO)

paper · pdf · doi:10.48550/arxiv.2411.02189

openalex publication_date 2024/11/04 · openalex created_date 2024/11/15 · openalex updated_date 2026/07/28

Abstract

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 vice-versa. 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.

Cited by

• Data-Efficient Multitask DAgger
• First Order Model-Based RL through Decoupled Backpropagation
• REBot: Reflexive Evasion Robot for Instantaneous Dynamic Obstacle Avoidance

Related

• Learning Deployable Locomotion Control via Differentiable Simulation
• Learning Quadruped Locomotion Using Differentiable Simulation
• Sim-to-Real: Learning Agile Locomotion For Quadruped Robots
• Residual Policy Learning for Perceptive Quadruped Control Using Differentiable Simulation
• Improving generalization of robot locomotion policies via Sharpness-Aware Reinforcement Learning