Driving · 2026-10-12

LeapBot-JAC: Joint Action Chains for World Action Models

Chenhao Zhu, Baibing Liu, Pei Liu, Rui Yang, Lang Zhang, Mingyue Feng, Yaonong Wang, Jun Ma

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

LeapBot-JAC addresses action-chunking policies that predict future controls but execute only a prefix before replanning. The authors propose LeapBot-JAC, which reorganizes a fixed action-token budget into a sparse reference chain and a dense execution chain within one shared action expert. Both chains are supervised by demonstrated controls and jointly denoised with aligned physical-time positions; directed attention lets execution tokens read reference tokens under the host’s visibility constraints.

Why it matters

Action-role organization is a design choice under a fixed token budget in world action models. The authors show that assigning distinct roles to reference vs. execution targets (without adding parameters or tokens and while preserving the host’s world-modeling process) improves success on LIBERO, LIBERO-Plus, and real-world deployment evaluations across two world-model hosts.

Method

Reallocate a fixed action-token budget into sparse future-control references and dense execution targets within one shared action expert (reference indices {0,3,...,45}, execution {0,...,15} with defaults).
Jointly denoise both chains with physical-time alignment; use directed role masks so execution tokens can read reference tokens while blocking direct reverse access, subject to host visibility constraints.
Execute only the execution prefix (first K controls) and replan; both chains participate in inference with no added action tokens or parameters.

Limitation

The authors state that their ablations assess direct visibility, positive reference-loss weights, and temporal strides, but do not identify reference semantics or prove abstract planning; sparse controls may act as auxiliary predictive targets, and blocked direct access may retain indirect communication. Isolating mechanisms would require additional matched controls, and no unperformed mechanism study is claimed as evidence.

Abstract (from arXiv)

Action-chunking policies predict future controls but execute only a prefix before replanning. How should a fixed action-token budget be organized to support execution? We propose LeapBot-JAC, which allocates sparse future-control references and dense execution targets within one shared action expert. Both chains are supervised by demonstrated controls and jointly denoised at aligned physical-time positions. Directed attention allows execution tokens to read reference tokens subject to the host's visibility constraints. JAC adds neither action tokens nor parameters and retains the host's world-modeling process; both chains participate in inference, while only the execution chain drives the robot. On LIBERO, JAC reaches 98.3% success with FastWAM and LeapBot-WA, up from 97.0% and 97.3%. On zero-shot LIBERO-Plus, their success rates improve from 53.5% to 70.4% and from 73.1% to 77.3%, respectively. RoboTwin and two real-world tasks provide additional deployment evaluations. Single-training-seed comparisons and FastWAM ablations support action-role organization as a practical design choice under a fixed action-token budget.

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First public release: 2026-10-09 · Indexed by Pub-AI: 2026-10-12 UTC

preprint