Event
VISTA: A Proactive Self-Exploration Framework for Self-Improving Embodied Agents
1 reports · 1 sources
Zhang Zhang and other authors propose and release the VISTA meta-agent framework, which changes the way the embodied agent harness is improved from "post-hoc replay reflection" to "proactive self-exploration." The framework maintains a harness graph that records the lineage of harness versions, allowing the meta-agent to retrieve historical hypotheses, reuse effective mechanisms, and avoid repeatedly failed directions; each round starts from retained versions and their execution evidence, autonomously generates new rollouts, probes failure and recovery behaviors, and then modifies the harness accordingly and re-evaluates it. The authors report that they evaluated VISTA on three benchmarks—LIBERO-Pro, RoboCasa, and RoboTwin—and achieved higher success rates with fewer simulator rollouts.
Report timeline
- VISTA: An Active Exploration Framework for Self-Improvement in Embodied Agents
VISTA is a meta-agent framework that replaces post-hoc replay reflection with active self-exploration to improve the harness of embodied agents. It maintains a harness graph that records the lineage of harness versions, enabling the meta-agent to retrieve past hypotheses, reuse effective mechanisms, and avoid repeating failed directions. Each round starts from a retained version and its execution evidence, autonomously generates new rollouts, probes failure and recovery behavior, and on that basis iteratively revises the harness and re-evaluates it.
arXiv cs.RO ·