IMPULSE: Push First, Then Predict in V-JEPA 2 Feature Space
IMPULSE: Push First, Then Predict in V-JEPA 2 Feature Space
This paper introduces IMPULSE, a technique that enables latent world models like V-JEPA 2 to infer object physical properties through brief calibration pushes. By using a Bayesian contact memory and a lightweight random-Fourier pose head, the method reduces 1-second feature-forecast error by approximately 11.6–12.8% on fresh objects in held-out evaluations across multiple training seeds. The approach demonstrates that motion-based pathways are necessary for physics to reach features, with gains scaling after contact onset and extending to out-of-range physics scenarios.
The study presents IMPULSE, a method that infers impulse responses in the frozen feature space of V-JEPA 2 to improve physics-aware prediction after brief calibration pushes; reported gains are relative reductions in 1-second feature-forecast error on held-out objects and seeds, with uncertainty intervals provided.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
gain multiplier over plain pose tokens
2.8–3.2
View original evidence
. On fresh objects, the Fourier head roughly triples the gain of a perfect memory over plain pose tokens (2.8-3.2 times). Raw push measurements fed through the same pathway give a similar gain, so the Bayesian summary is not required; it condenses the trials into eight interpretable numbers. The gain is absent at contact onset and grows afterwards.
Open source S5
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
ure space the same ability. A Bayesian contact memory summarizes three short calibration pushes, a predicted object pose carries it into the predictor, and a lightweight random-Fourier pose head lets the predictor use the pose. On 512 fresh objects, in a pre-specified held-out confirmation with models sealed before the data existed, the memory lowers the 1-second feature-forecast error by 11.6-12.8% relative to the s
Open source S3
ame predictor without experience, for all 8 training seeds, with a 95% seed-and-object interval excluding zero. The gain also holds for 128 objects with out-of-range physics (11.7-14.0%), and the full extension beats a plain predictor by 15.5-16.6%. An information-ladder argument, exact for ideal predictors, shows that physics reaches the features only through motion, so we measure how much the pathway lets pose help
Open source S4
. On fresh objects, the Fourier head roughly triples the gain of a perfect memory over plain pose tokens (2.8-3.2 times). Raw push measurements fed through the same pathway give a similar gain, so the Bayesian summary is not required; it condenses the trials into eight interpretable numbers. The gain is absent at contact onset and grows afterwards.
Open source S5
Source：arXiv Robotics — research abstracts · arxiv.org