From Reaction to Anticipation: Proactive Failure Recovery through Agentic Task Graph for Robotic Manipulation

Venue: RSS 2026 (Sydney, Jul 13–17) · Session: Planning · paper #180 Authors: Sheng Xu, Ruixing Jin, Huayi Zhou, Bo Yue, Guanren Qiao, Yueci Deng, Yunxin Tai, Kui Jia, Guiliang Liu arXiv: 2605.11951 · program page

Summary compiled from the arXiv paper (v1, 12 May 2026); all numbers quoted from the paper. Trend context: RSS 2026 survey.

Fig. 1: On a dual-arm pour-water task, reactive baselines either backtrack to the last keyframe (losing progress) or re-plan via a VLM/LLM (high latency, high cost). AgentChord instead pre-generates recovery branches — recovery nodes, edges and compiled monitors — for anticipated errors (cup shifted/tilted, bottle shifted/not fully grasped), enabling instant online recovery; the bottom strip shows the recovery being executed.

Problem

Task failures are inevitable in dynamic, unstructured manipulation. Conventional detect–reason–recover pipelines react only after a failure, incurring high latency and limited robustness because reasoning and planning happen online, after the fact.

Method

AgentChord models a manipulation task as a directed, recovery-augmented graph. Before execution, the graph is enriched with anticipatory recovery branches specifying context-aware corrective behaviors. Execution is a choreography of specialized agents: a composer that structures the nominal task graph, an arranger that augments it with recovery branches and compiles execution, and a conductor that orchestrates recovery. Low-latency compiled monitors detect deviations and trigger the pre-compiled recoveries without re-planning.

Results

On three simulated bimanual tasks with disturbances, AgentChord attains the best average success rate of 99.2% versus CaM 97.5%, ReKep 92.5%, DRM 92.5% and an imitation baseline (IM) 79.2%, while achieving the lowest average execution time of 41.5 s (vs ReKep 54.4 s, CaM 78.1 s). On six real-world tasks with disturbances it averages 77.5% success versus CaM 72.5%, ReKep 65.0%, DRM 66.7% and IM 59.2%, with the lowest average execution time of 92.2 s.

Significance

Shifts failure handling from reaction to anticipation, improving reliability and efficiency on long-horizon bimanual manipulation. Connects to Review-System-0-1-2 and Review-VLA-Memory.

← Back to RSS 2026 survey · RSS-2026-Papers · Home