SG-ODOM’s developers report approximately 25% lower translational relative pose error—error in estimated changes in position—than ANYbotics’ onboard estimator across seven outdoor ANYmal D missions in the GrandTour dataset. A foot hitting the ground hasn’t necessarily earned the motion estimator’s trust. Source
Frank Dellaert shared the framework on November 18, describing a collaboration with Oxford, SNU and Pisa. SG-ODOM estimates a legged robot’s movement using its own inertial and joint-related measurements, known as proprioception. This is motion estimation, not a new walking controller. Source Source
Touchdown isn’t the same as stable footing
According to the project page, impact and load transfer mean a newly planted foot doesn’t necessarily provide a stationary reference. SG-ODOM estimates vertical ground reaction force from joint torques, then accepts a contact only after the load is sufficient and force variation stays limited over a validation interval. Source
Each validated touchdown immediately triggers a correction to the estimate. Between corrections, an inertial measurement unit, or IMU, supplies the measurements used to track motion. There’s also a timeout: if no correction arrives within a maximum interval, the system applies an update rather than continuing on inertial measurements alone indefinitely. Source
SG-ODOM still processes all contact measurements. Closely spaced, strongly correlated measurements from the same contact phase no longer trigger repeated corrections. Source
Tests on Spot, ANYmal D and Unitree G1
The team reports testing on Boston Dynamics Spot, ANYbotics ANYmal D and the Unitree G1 humanoid. It says the same estimation back end runs on quadrupeds and humanoids, with only a small robot-specific front end changing. Four legs or two, the back end keeps its job. Source
That approximately 25% reduction is specific to the ANYbotics comparison across those seven ANYmal D missions. The project page doesn’t report the same gain across all three platforms. Source
The project page links an arXiv preprint and a code repository. Its summary gives no absolute errors or RPE evaluation interval. Better motion estimates aren’t proof of better walking. Source
Comments