Zero-shot sim-to-real transfer using Siamese-Q-Based reinforcement learning

To address real world decision problems in reinforcement learning, it is common to train a policy in a simulator first for safety. Unfortunately, the sim-real gap hinders effective simulation-to-real transfer without substantial training data. However, collecting real samples of complex tasks is often impractical, and the sample inefficiency of reinforcement learning exacerbates the simulation-to-real problem, even with online interaction or data. Representation learning can improve sample efficiency while keeping generalization by projecting high-dimensional inputs into low-dimensional representations. However, whether trained independently or simultaneously with reinforcement learning, representation learning remains a separate auxiliary task, lacking task-related features and generalization for simulation-to-real transfer. This paper proposes Siamese-Q, a new representation learning method employing Siamese networks and zero-shot simulation-to-real transfer, which narrows the distance between inputs with the same semantics in the latent space with respect to Q values. This allows us to fuse task-related information into the representation and improve the generalization of the policy. Evaluation in virtual and real autonomous vehicle scenarios demonstrates substantial improvements of 19.5% and 94.2% respectively over conventional representation learning, without requiring any real-world observations or on-policy interaction, and enabling reinforcement learning policies trained in simulations transfer to reality.