TRACE: New AI Odometry for Legged Robots Under Unreliable Contact
A novel proprioceptive odometry estimator for legged robots, named TRACE (Tokenized Robust Attention for Contact-Aware Estimation), has been developed by researchers to function effectively in unreliable contact scenarios. This system predicts relative displacement, relative rotation, and body-frame velocity using recent inertial and joint data from onboard sensors. Its foot-aware cross-attention module dynamically adjusts the importance of IMU and leg-specific kinematic tokens, eliminating the need for predefined contact or slip thresholds, thus improving robustness. The training process includes direct supervision and two auxiliary losses inspired by physics to ensure kinematic consistency and effective leg information utilization. To enhance sim-to-real transfer and reduce policy-specific overfitting, the training in simulation features policy randomization, followed by partial fine-tuning in real-world conditions. The research is published on arXiv with the identifier 2608.05975.
Key facts
- TRACE stands for Tokenized Robust Attention for Contact-Aware Estimation.
- It is an end-to-end learned proprioceptive odometry estimator for legged robots.
- It predicts relative displacement, relative rotation, and body-frame velocity.
- Uses a foot-aware cross-attention module to weight IMU and leg-wise kinematic tokens.
- No manually defined contact or slip thresholds are required.
- Trained with direct supervision and two physics-inspired auxiliary losses.
- Simulation training includes policy randomization and partial real-world fine-tuning.
- Paper available on arXiv: 2608.05975.
Entities
—