TL;DR
SG-ODOM is a new proprioceptive odometry framework for legged robots that improves state estimation by being picky about when it listens to leg sensors. By using a "stability gate" to ignore noisy impact transients and an event-driven update schedule to avoid redundant data, it reduces translation errors by up to 25% across quadrupedal and humanoid platforms.
What Problem Does This Paper Address?
Legged robots usually estimate their movement by fusing high-speed Inertial Measurement Unit (IMU) data with leg kinematics (encoder data). The standard assumption is that when a foot is on the ground, it is a stationary reference point.
However, reality is messy. When a foot first hits the ground, there is a transient phase of impact, bouncing, and load transfer where the foot isn't actually stable. If an estimator treats these noisy moments as "stationary," it introduces errors. Furthermore, most systems update at the full rate of the leg encoders (often 400Hz+). Because measurements within a single footstep are highly correlated, updating this frequently doesn't just add redundant information—it can actually amplify errors caused by slight foot slips or residual motion, leading to "drift" in the robot's perceived position.
Methods
The authors propose SG-ODOM, which decouples high-rate sensing from the timing of estimator updates through two main mechanisms:
- Stability-Gated Contact Classification: Instead of trusting a contact the moment a force sensor triggers, SG-ODOM waits. A contact is only "validated" for estimation if the Ground Reaction Force (GRF) exceeds a threshold and the variation in that force remains low for a specific window of time. This ensures the estimator only uses the foot as a reference once the impact transients have settled.
- Event-Driven Updates: Rather than updating at a fixed high frequency, the system triggers kinematic corrections based on two events:
- Validated Touchdown: An immediate update occurs the moment a foot is deemed stable.
- Timeout ($\Delta t_{max}$): If a foot stays in contact for a long time, updates are triggered at a lower, controlled rate to prevent IMU drift without over-incorporating redundant kinematic data.
The framework is modular, featuring a robot-specific "front end" and a shared "back end" that supports two common estimator types: an Invariant Extended Kalman Filter (InEKF) and a Fixed-Lag Smoother.
Datasets
The researchers evaluated SG-ODOM across three distinct legged platforms and several environments:
- GrandTour Dataset: Used for ANYmal D (quadruped) in diverse outdoor terrains (forest, snow, industrial).
- GaRLILEO Dataset: Used for Boston Dynamics Spot (quadruped) in indoor/outdoor sequences including stairs and overpasses.
- Self-Collected Dataset: Used for the Unitree G1 (humanoid) on paved courtyards and slopes to test generalization to different morphologies.
Key Insights and Ideas
- More Data $\neq$ Better Estimates: A counter-intuitive finding is that updating the estimator at the maximum sensor rate often increases error. Lowering the update frequency (e.g., to 10Hz or 20Hz during a stance phase) allows the IMU to handle short-term motion while the kinematics provide sparse, high-quality "anchors."
- The "Impact Gap": By explicitly waiting for the force to stabilize after touchdown, the system avoids the most non-linear and noisy part of the gait cycle. The "effective delay" created by the stability gate isn't fixed; it adapts to how quickly the foot settles on different terrains.
- Platform Agnosticism: The logic holds up across different robots. Whether it's a heavy quadruped on snow or a lightweight humanoid on concrete, the principle of "stability-gated" updates consistently outperforms standard fixed-rate methods.
Key Results
- Accuracy Gains: SG-ODOM-FL (the fixed-lag smoother version) achieved a 25% reduction in average translational Relative Pose Error (RPE) compared to the onboard ANYbotics estimator on the GrandTour dataset.
- Robustness: On the Spot robot, SG-ODOM outperformed the well-known Pronto estimator, particularly in challenging sequences like stair descents where impact noise is high.
- Optimal Update Window: The ablation studies showed that for most robots, an update interval ($\Delta t_{max}$) of 0.1 seconds (10 Hz) provided the best balance between constraining IMU drift and avoiding kinematic noise.
Limitations and Future Work
- IMU Calibration: The authors noted that they used empirical tuning for IMU biases rather than a formal Allan-deviation analysis, as dedicated calibration data wasn't available for all platforms.
- Proprioception Only: While SG-ODOM is highly effective, it is still a proprioceptive-only system. It cannot correct for long-term global drift that exteroceptive sensors (like LiDAR or Cameras) could handle.
- Future Directions: The authors suggest that future work could involve adaptive contact-validation criteria (e.g., changing the stability thresholds based on terrain type) and tighter integration with vision or LiDAR-based SLAM systems.
For those interested in the implementation, the source code is available on GitHub.
Summary updated