Zhang Zhang et al. (7 authors) · 2026-10-26
VISTA: Active Exploration for Self-Improvement in Embodied Agents
“Experimental results demonstrate that VISTA progressively improves task success and is more rollout-efficient than static reflection, achieving higher success rates with fewer simulator rollouts.”
- Who
- Zhang Zhang et al. (7 authors)
- Date
- 2026-10-26
- Evidence
- Very low certainty Starts at Low: the developer’s own numbers.
Very lowLowModerateHigh
Low ↓ risk of bias ↓ imprecise ↓ indirect measure → Very low
- ↓ risk of bias. The reported rates belong to the harness kept by best-so-far selection on the same tasks it was developed on, one harness per task. The paper reports no separate test after selection. With early stopping on, "a confirmation iteration re-evaluates that package without another development attempt", on the same tasks. “Likewise, the non-decreasing champion curves follow from best-so-far selection and do not imply that every proposed revision improves performance.”
- ↓ imprecise. Each LIBERO-Pro and RoboCasa365 task is scored on 16 episodes, eight scene seeds with two repetitions; RoboTwin uses four scenes with two repetitions. The paper gives no confidence intervals and no repeated development runs. “Solid cells report candidate successes out of 16 episodes.”
- ↓ indirect measure. The comparison with static reflection in the quote rests on a passive variant run on four LIBERO-Goal tasks for two rounds, and that table counts Meta tokens, not simulator rollouts. “Token is the mean cumulative Meta tokens per task after each round.”
- Replication
- No rerun found. Code is linked from the paper. We found no rerun by another group. Searched up to 2026-10-28. all claims
- Checked
- 2026-10-28, against the source page
- Links
- Source Paper Code
Also on the page
- A meta-agent, Codex with GPT-6, revises the harness of a robot agent: its instructions, executable skills, policy hooks, stored procedural knowledge and tool bindings. The planner (GPT-5.6 Sol), the VLA policy, the capability interface and the evaluation rule stay fixed. Every frozen version is a Git commit in a harness graph, and the meta-agent can build on an earlier or rejected version.
- Against the initial harness, mean success goes from 17.5% to 95.0% on eight LIBERO-Pro settings, where HarnessVLA scores 61.1%. On RoboCasa365 it goes from 16.4% to 24.2% on Composite Seen and from 4.7% to 13.3% on Composite Unseen, where RLDX-1 scores 17.4% and 6.8%. On RoboTwin it goes from 48.5% to 74.0%, where HarnessVLA scores 58.0%.
- A harness is developed and selected for each task, and the reported rates are the scores of the retained version on those tasks; no tasks are kept out of development. Composite Unseen is a RoboCasa365 category, and VISTA developed harnesses on those tasks too. Against a passive variant on four LIBERO-Goal tasks, active exploration gains 85.9 points after two rounds against 56.3, using 91.6M Meta tokens against 137.7M.
- The authors' limits: "Physical real-robot validation has not yet been performed."
In the self-improving AI catalog
- VISTA: what it changes, whether the change lasts, and how it was tested
How it got here
Earlier work in this catalog that VISTA cites, oldest first. Under each: where that work sits on the map, then where VISTA sits (changes in bold), and what VISTA's own paper says about it. Its whole family on the citation map