DASH-ACT: Demonstration-Adaptive Step-Horizon Action Chunking for Manipulation
Action Chunking Transformer (ACT) policies predict future action sequences, but their execution rules determine how quickly new observations can influence control. We present DASH-ACT, a supervised ACT extension combining motion-derived heuristic horizon labels, multihorizon prefix reconstruction, and a learned commitment selector. The policy predicts 20 actions and commits 5, 10, or 20 before replanning. On three LIBERO-10 tasks, mean success rates across three training seeds increase from 70.7% to 78.0%, 46.0% to 67.3%, and 68.0% to 91.3% relative to the evaluated fixed-20-step ACT baseline. These differences are descriptive and do not reach training-seed-level significance after correction across tasks. Separate execution controls reuse the trained DASH-ACT predictor with fixed horizons. Their post-hoc best per-task settings have slightly higher mean success than adaptive selection, leaving the benefit of the learned selector unresolved. In an independently trained physical eraser pick-and-place evaluation with lateral human perturbations, DASH-ACT succeeds in 44 of 50 trials, compared with 17 of 50 for ACT. The results identify queue commitment as a useful design variable within ACT and motivate direct label validation and component attribution.
Authors
- Yibo Cao (ORCID: https://orcid.org/0000-0002-8140-0043)
- Jiahong Huang (ORCID: https://orcid.org/0009-0009-9049-0173)
- Zijun Yin
Institutions
- South China Normal University (CN)
Publication Details
- Journal
- Sensors
- Published
- 2026-11-17
- DOI
- https://doi.org/10.3390/s26227239
- Primary Topic
- Robot Manipulation and Learning
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- South China Normal University