VISTA from Tsinghua and NUS lifts LIBERO-Pro robot success from 17.5% to 95% without weight updates

量子位 · wechat · 2026-11-07

Researchers from Tsinghua University and NUS open-sourced VISTA, an embodied agent framework that lets robots improve their own skills through self-directed experiments and code development — with model weights fully frozen.

How it works

An executor agent completes tasks via a "Harness" (code layer organizing instructions, skills, tool calls, and feedback); a Meta-agent acts as the developer, reviewing code and run logs, writing or revising skills, and entering simulation when existing evidence is insufficient;
Candidate Harnesses are frozen for evaluation and adopted only if they beat the current version; code versions, lineage, and evidence live in a Harness graph enabling branch exploration;
Improvements persist across rounds: on a cream-cheese shelving task, an initially 0/16 Harness reached a reproducible 16/16 after two development rounds.

Results: LIBERO-Pro overall success rose from 17.5% to 95.0%; LIBERO-T/S from 22.8%/12.2% to 96.6%/93.4%; RoboTwin's 50 tasks from 48.5% to 74.0%; RoboCasa Seen/Unseen from 16.4%/4.7% to 24.2%/13.3%.

The team frames this as a step toward recursive self-evolution for embodied agents: robot capabilities keep growing after deployment through experience-driven skill development.

Original post →

VISTA from Tsinghua and NUS lifts LIBERO-Pro robot success from 17.5% to 95% without weight updates

More from Embodied