SplineWAM: Adaptive Action Horizons for World Action Models via B-Spline Representations
This preprint introduces SplineWAM, which replaces the fixed-length action chunk of a world action model with a fixed-size window of cubic B-spline knot times and control points, letting each prediction adapt its temporal horizon and resolution. On LIBERO-Plus and RoboCasa it improves success rate by 8.2 and 4.4 points over an action-chunking WAM while issuing 22% and 26% fewer policy calls per episode; on three real bimanual tasks it matches or beats the baseline while decoding 1.2–1.6× more executed motion per call. This matters for cloud-served robot fleets because fewer calls per task directly increase throughput.
Can a world action model predict action trajectories with adaptive temporal horizon and resolution from a fixed-size B-spline parameter budget, and can that representation be deployed asynchronously on real robots?
World action models typically output a fixed number of actions sampled on a uniform grid, so they spend the same inference budget on smooth free-space motion and on contact-rich manipulation; this creates redundant policy calls in easy phases and stale actions in precise phases, limiting both task success and cloud inference throughput.
Prior WAMs used action chunking after Zhao et al. 2023, predicting fixed-length uniform action sequences. Spline tokenization variants such as BEAST keep a fixed represented duration, and other adaptive-horizon methods adjust chunk size from auxiliary signals at inference, but none lets the represented duration vary directly from the predicted action parameters.
SplineWAM keeps the Fast-WAM backbone but replaces its action branch with a fixed-size window of cubic B-spline knot times and control points. Demonstrations are fit offline under an error tolerance so knots concentrate where the trajectory is complex, and video supervision is sampled at the fitted knot times rather than a uniform grid. For asynchronous execution, Jacobian-pullback real-time chunking (JP-RTC) enforces continuity on decoded actions and corrects spline parameters through the decoder's local Jacobian.
On LIBERO-Plus (ε=0.01), SplineWAM-Knot reaches 74.5% success vs 66.3% for action chunking (+8.2 points) with 26.2 vs 33.6 calls per episode (-22%) and 8.41 vs 8.00 executed steps per call. On RoboCasa (ε=0.08), it reaches 73.5% vs 69.1% (+4.4 points), 16.3 vs 21.9 calls per episode (-26%), and 18.18 vs 16.00 steps per call. BEAST, which uses fixed-duration spline tokenization, scores 66.6% on LIBERO-Plus and 65.0% on RoboCasa, not improving. Removing the video auxiliary term from Naive B-Spline costs 10.7 points on LIBERO-Plus and 14.5 points on RoboCasa. In ablation sweeps, matching the action-chunking baseline's best accuracy requires 2.2× fewer policy calls per episode on RoboCasa and 2.5× fewer on LIBERO-Plus. Real-robot evaluation over 30 trials per task per method under asynchronous execution: JP-RTC scores 23.3 SR/54.4 PG on Arrange Bookshelf vs 20.0/52.2 baseline; 26.7/57.5 on Charge Earphone vs 26.7/73.3 (same SR, lower progress at insertion); 60.0/74.2 on Clean Whiteboard vs 50.0/67.5 (+10.0 SR, +6.7 PG). Executed chunk lengths are 38.6, 45.7, and 52.4 control steps vs the fixed 32.0 baseline, i.e. 1.2–1.6× more executed motion per call. Naive RTC underperforms the baseline on all three tasks despite longer chunks.
The authors note that the continuous representation is less reliable at precise socket insertion in Charge Earphone, where the action-chunking baseline leads in progress. Real-robot validation is limited to three bimanual tasks and 30 trials per method per task; simulation covers two suites. The fitting tolerance and executed row budget are chosen per suite, demonstrations must be fit offline, and the paper is a preprint with no stated peer review or code release. JP-RTC relies on local Jacobian and regularization approximations, and fleet-scale serving was not tested.
Cloud robotics providers and robot OEMs deploying world action models or VLA policies at fleet scale could use SplineWAM to reduce inference calls per task and improve success without retraining the video backbone. Because it is compatible with existing WAM backbones, adoption could begin within 1–3 years if JP-RTC is validated across more contact-rich and high-precision tasks, the per-suite tuning is automated, and the method is integrated into cloud serving stacks.
The full text is not republished here because the paper's license does not allow it. Read the original on arXiv.