In autonomous robotic operations, decisions may need to be made based on evidence that is no longer visible. We investigate extbf{delayed-evidence} tasks, where an early cue disappears before a later decision point, necessitating different actions for visually similar observations. In these contexts, the current observation is insufficient for control.
We introduce extbf{Trajectory-Routed Causal Evidence} (TRACE), a memory framework for visuomotor imitation policies. TRACE stores task-relevant visual and robot-state evidence, such as object identity, target choice, or route-dependent states, in a fixed-size latent memory that remains bounded over long episodes. Rather than indexing memory by raw time or manually provided task labels, TRACE employs extbf{path signatures}: compact, order-sensitive features of the executed robot-state trajectory. These signatures do not store the visual cue itself; instead, they provide trajectory-conditioned keys for writing and retrieving the evidence stored when the cue was visible.
When the robot later encounters an ambiguous observation, the policy conditions on TRACE memory to recover the missing context and choose the correct branch. TRACE attaches through lightweight adapters to policies, without altering the policy backbone, action head, or imitation objective. Across real-world long-horizon manipulation tasks with visually ambiguous branch points, TRACE improves branch selection and task success over alternative baselines, including short-history and recurrent memory.
Blogger's Review: TRACE introduces a novel approach to memory mechanisms in robotics, particularly in dynamic environments where decisions must adapt to changing contexts. Its ability to handle visual ambiguity significantly enhances task success rates, showcasing the importance of contextual memory in complex robotic tasks. This framework merits further exploration and application in advanced robotic systems.