Event

TaskAnchor Uses Execution History to Alleviate VLA Task-State Confusion

1 reports · 1 sources

A research team including DexForce from Cross-Dimensional Intelligence published a paper proposing the TaskAnchor method, which uses a lightweight adapter to connect execution history to the original language and vision interfaces of pretrained VLAs (π₀.₅, XVVA), alleviating task-state confusion in long-horizon manipulation. The method adopts a dual-path design: a normalized task-state coordinate describes the semantic execution stage, and a visual residual based on context-adaptive test-time training (CA-TTT) preserves historical details. Under the single-task training setting of RMBench, TaskAnchor raised the average success rate of π₀.₅ from 14% to 79%, with an additional inference overhead of 2.08 milliseconds per action chunk and no additional online fine-tuning during testing. The paper notes that this latency corresponds only to model inference and is not equivalent to the end-to-end control latency of a complete robotic system.