Francesco Iotti*, Chiyun Noh*, Franco Angelini, Manolo Garabini, Ayoung Kim, Maurice Fallon, Frank Dellaert

* Equal contribution

Arxiv
Arxiv
YouTube
YouTube
Code
Code

SG-ODOM deployed on Boston Dynamics Spot, ANYbotics ANYmal D and Unitree G1
SG-ODOM has been tested on three legged robots: Boston Dynamics Spot (left), ANYbotics ANYmal D (center) and the Unitree G1 humanoid (right).

SG-ODOM is a proprioceptive odometry framework for legged robots that fuses IMU propagation with leg kinematics, using stability-gated contact selection to decide when kinematic corrections are applied. Physical touchdown does not necessarily provide a stationary reference, particularly during impact and load transfer, so a contact is accepted for state estimation only once the foot reaches a sufficiently stable support condition. Each validated touchdown triggers an immediate estimator correction, while timeout updates bound the maximum duration of IMU-only propagation.

The same back end runs on quadrupedal and humanoid robots, and is implemented in GTSAM both as an invariant extended Kalman filter (InEKF) and as a fixed-lag smoother. Only a small robot-specific front-end changes between platforms. SG-ODOM reduces translational Relative Pose Error (RPE) by approximately 25% compared to the onboard ANYbotics estimator across seven outdoor missions of the GrandTour dataset.

Tested Platforms

Robot Type Data
Boston Dynamics Spot Quadruped Four indoor and outdoor sequences from the GaRLILEO dataset
ANYbotics ANYmal D Quadruped Seven outdoor missions from the GrandTour dataset
Unitree G1 Humanoid Two self-collected outdoor sequences
GrandTour: ANYmal on PIL-2
GrandTour: ANYmal on PIL-2 (vs. ANYbotics estimator)
GaRLILEO: Spot on DOWNSTAIR
GaRLILEO: Spot on DOWNSTAIR (vs. MUSE)
Unitree G1 on SLOPE
Unitree G1 on SLOPE (vs. Pronto)

SG-ODOM-FL (blue) tracks the reference trajectory more closely than the baselines (red), with noticeably less vertical drift on stairs and slopes.

Method Overview

SG-ODOM framework overview

A robot-specific front end converts IMU and joint-encoder measurements into common inputs: IMU measurements, foot positions from forward kinematics, and contact information. The contact classifier estimates the vertical ground reaction force from joint torques and accepts a contact only after the load is sufficient and the force variation stays limited for a validation interval, rejecting the transient that follows impact. A shared InEKF or fixed-lag smoother back end propagates the IMU at full rate and applies kinematic corrections only at validated touchdowns, or when the time since the last correction reaches a maximum dead-reckoning interval.

Event-based contact updates for a trotting quadruped

Because every touchdown resets the dead-reckoning interval, corrections follow the sequence of contact events rather than a fixed sampling grid. All contact measurements are still processed, but closely spaced, strongly correlated measurements from the same contact phase no longer trigger repeated corrections.

Citation

@article{iotti2026sgodom,
  title={{SG-ODOM}: Event-Driven Proprioceptive State Estimation with Stability-Gated Contact Updates for Legged Robots},
  author={Iotti, Francesco and Noh, Chiyun and Angelini, Franco and Garabini, Manolo and Kim, Ayoung and Fallon, Maurice and Dellaert, Frank},
  journal={arXiv preprint arXiv:2611.20710},
  year={2026}
}