Reproduce “AutoOdom: Learning Auto-regressive Proprioceptive Odometry for Legged Locomotion” on OpenResearch

About this paper

Changsheng Luo, Yushi Wang, Wenhan Cai, Mingguo Zhao

AutoOdom is a proprioceptive odometry system for legged robots that employs a two-stage training paradigm and an autoregressive enhancement mechanism. This method effectively bridges the sim-to-real gap by combining large-scale simulation data with limited real-world fine-tuning, leading to a 59.2% improvement in Absolute Trajectory Error (Umeyama alignment) over prior state-of-the-art on a humanoid robot.

Key insights

• A two-stage training strategy effectively leverages abundant simulation data for initial learning of complex dynamics and then adapts to real-world conditions with minimal real-world data, successfully bridging the sim-to-real gap.
• The autoregressive feedback loop is crucial for the model to learn robust predictions that inherently account for accumulated errors and enhance resilience against sensor noise in dynamic environments.
• Optimal sensor input configurations vary between training stages; specifically, excluding simulated linear acceleration in Stage 1 but reintroducing real IMU acceleration in Stage 2 improves overall performance.

Summary by alphaXiv.