EN
RESPLE2 predicts robot motion from inertial data
RESPLE2: Inertial-Informed Recursive Non-Uniform Spline Estimation for Multi-Sensor Odometry with Global Refinement · arXiv:2610.37602
In plain words
RESPLE2 is a method for estimating how a robot moves by combining readings from sensors such as cameras, LiDAR and inertial sensors. It represents motion as a changing curve and can use inertial readings to predict the curve’s next point, then update it with other sensor readings.
In nine enclosed-space HOPPER sequences, this inertial prediction lowered position error compared with predicting without inertial data. That result applies to those sequences.
The research
When robot sensors arrive at different times, odometry must combine them while tracking rapid motion.
It remains unclear whether a recursive system can adapt its motion representation and make useful predictions from inertial readings, rather than relying on fixed intervals and approximate state initialization.
RESPLE2 adapts a continuous-time motion curve and uses inertial readings to predict new points as sensor measurements arrive.
RESPLE2 represents six-degree-of-freedom robot motion with a non-uniform spline, placing new points according to motion and measurement availability.
At each boundary it predicts the newest control point from inertial readings, then updates the trajectory with sensor measurements at their recorded times.
On nine enclosed-space HOPPER sequences, the authors’ ablation found 39%–76% lower position error than prediction without inertial data, with knot intervals and measurement batches held constant. This supports a benefit from inertial prediction in that test.
Separately, the leg-kinematic variant had the lowest reported position error on eight JEPLO sequences.
The HOPPER prediction comparison covers nine enclosed-space sequences, and the global-refinement result is from one loop closure in one HANDI trajectory.
These evaluations do not establish performance across untested platforms or sensor conditions; the authors identify robustness under sensor degeneracy as future work.
What the evidence shows
- The asynchronous sensing problem
RESPLE2 is a recursive multi-sensor odometry framework that estimates a robot’s motion over time. The research addresses a challenge in combining readings that arrive at different times: fixed-time motion states and interpolation can complicate sensor fusion and handling motion distortion.
The framework instead represents motion with a continuous-time curve and can incorporate measurements at their recorded times.
- Predicting the next motion point
The curve is a non-uniform B-spline, which uses control points that need not be evenly spaced. RESPLE2 places new points according to motion and measurement availability.
At a clamped boundary, the newest control point lies on the represented trajectory, so the system can predict it by integrating inertial readings. A reset then preserves key motion quantities at the previous endpoint as the curve extends.
- What the prediction test shows
The authors tested the prediction step in an ablation on nine enclosed-space HOPPER sequences recorded during aggressive motion. They compared IMU-informed prediction with a condition that copied the latest position and orientation-increment control points.
Knot intervals and measurement batches were held constant. The IMU-informed condition reduced position APE, the root mean square position difference from ground truth, by approximately 39% to 76% across those sequences.
- Evidence on quadruped runs
A separate comparison used eight JEPLO sequences of a quadruped making dynamic maneuvers over varied terrain. Among the systems and configurations listed by the authors, RESPLE2’s leg-kinematic variant had the lowest reported position APE on every sequence.
This supports the variant’s performance on those tests, while the comparison does not establish that all systems used matched sensor inputs or identical procedures.
- A separate trajectory correction
RESPLE2 also offers global trajectory refinement after detecting a loop closure. In one HANDI handheld trajectory, the first detected loop closure was followed by a reported reduction of up to nearly 2 meters in position error along the trajectory, while angular and linear velocities were preserved.
That result demonstrates the correction in one event; the evidence remains bounded to the reported datasets, platforms, sequences and sensor setups.
Deep
- Original paper authors
- Ziyu Cao, William Talbot, Song Li, Pakpong Chirarattananon, Fredrik Gustafsson
- Submitted to arXiv
- License
- arXiv non-exclusive