New self-supervised framework boosts legged robot odometry without force sensors
Robot legs now track steps using only joint sensors, beating force-based methods.
Legged robot odometry—estimating a robot's position based on leg movement—relies on accurately detecting when a foot is in contact with the ground (stance phase) versus when it's swinging. Typically, this requires force sensors at the foot tip, which are expensive, fragile, and often miss phenomena like slippage.
Now, researchers Emre Girgin and Cagri Kilic introduce a self-supervised representation learning framework that uses only the standard joint encoder sensors already present in most legged robots. It learns to detect contact probabilistically, without labeled data or augmented hardware. In experiments, the framework outperformed supervised methods that require force sensors, as well as traditional probabilistic baselines. The team has released the code publicly, enabling immediate adoption by the robotics community for more robust, low-cost odometry.
- Uses only joint encoders—no force sensors needed—reducing hardware costs and complexity.
- Self-supervised learning eliminates the need for manual labeling of contact phases.
- Handles ground slippage better than force-based detectors, improving odometry accuracy in real-world conditions.
Why It Matters
More reliable robot locomotion navigation with standard sensors, paving way for affordable, robust legged robots.