This is the authors' abstract. We don't add key points for this paper.
Task and Motion Planning (TAMP) couples symbolic decision making with continuous robot-motion generation, yielding a highly multimodal hybrid state space. Recent unified-state TAMP planners, such as Task and Motion Informed Trees (TMIT*), offer anytime performance with almost-surely asymptotically optimal (ASAO) guarantees. However, planning can stall near sparse mode transitions and become dominated by repeated low-level geometric computations as the number of reachable modes grows. This paper presents Cross-mode Task and Motion Informed Trees (CmTMIT*), an acceleration of TMIT* that addresses both issues without changing the underlying ASAO search framework. CmTMIT* strengthens inter-mode connectivity by locally refining samples around validated transitions. It also reduces redundancy by sharing mode-invariant robot configurations and reusing expensive precondition projections across modes. CmTMIT* further amortizes k-nearest-neighbour (kNN) computation via a cross-mode cache built on the shared sample stream. Experiments on clutter clearing, obstacle-augmented clutter clearing, and shelf rearrangement benchmarks demonstrate consistent improvements over TMIT*. CmTMIT* achieves higher success rates under fixed timeouts, shorter initial-solution times, and improved anytime cost reduction, particularly as the number of objects and reachable modes increases.
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
Fukuda et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: