VISTA: Active Exploration for Self-Improvement in Embodied Agents
Zhang Zhang, Shengjun Zhang, Simin Huang, Min Chen, Chong Xia, Xin Fei, Yueqi Duan
cs.RO
2026-10-26
VISTA freezes weights and has a coding meta-agent rewrite the robot harness after live probes. LIBERO-Pro success rises from 17.5% to 95.0%, and RoboTwin from 48.5% to 74.0%.
On long-horizon manipulation, an end-to-end VLA (a vision-language-action model that maps images and instructions straight to motor commands) stacks small execution errors along the trajectory. Once a language model sits above it as a planner, finishing the task depends on the harness: the executable control surface around the frozen models, covering instructions, skill code, tool bindings, and recovery.
Most agent systems treat a finished rollout, one full interaction trace, as a static log and edit the harness afterwards. Those logs often lack the observation or the intervention that would test a failure hypothesis. The edit never goes back into the scene.
VISTA hands the rewrite to a coding meta-agent that has to collect that evidence in simulation. Policy and planner weights stay fixed.
The online agent can call a language-model planner, one VLA, and a public sensing and control interface. Only the composition is searched. A harness packages model-facing instructions, executable skills, request and feedback hooks, procedural memory, and tool bindings. Skills chain perception and control; hooks decide when they run.
Versions form a directed acyclic graph stored as Git commits, so each parent link is a lineage. Search starts from the current champion and may also reopen a rejected version to reuse a mechanism that still works. Failed proposals and their diagnostics stay on the graph.
A round does not end when the log has been read. Inside the budget the meta-agent chooses the next move: inspect history, edit code, open a private simulator from a frozen snapshot for live diagnosis, or run an autonomous test of an intermediate build. Live probes can read state and call sensing or control to check one hypothesis. Those traces are development evidence. They do not count toward the official score.
A submitted candidate is frozen. The host re-evaluates it on the declared development matrix with fresh episodes and no meta-agent in the loop. It replaces the champion only when empirical success is strictly higher. Ties keep the incumbent. With early stopping, the same package must post consecutive perfect formal scores; a confirmation round retests and does not edit.
The inner loop gathers evidence and rewrites a candidate. The outer loop keeps a version only on the frozen score.
Weights stay frozen on all three benchmarks. The low-level policies are π0.5 on LIBERO-Pro, RLDX-1 on RoboCasa365, and LingBot-VLA on RoboTwin. The planner is GPT-5.6 Sol. The meta-agent is Codex with GPT-6. On LIBERO-Pro the initial harness H0 exposes only the frozen VLA and a thin observation interface. On the other two benchmarks H0 is adapted from HarnessVLA, keeping its prompts, hooks, tools, and memory. Each task keeps its own evolved package, then that package is frozen for evaluation. A LIBERO suite has 10 tasks, 8 scene seeds, and 2 repetitions per seed.
| Benchmark | Reference | H0 | VISTA |
| LIBERO-Pro, mean of 8 settings | HarnessVLA 61.1%, π0.5 11.0% | 17.5% | 95.0% (+77.5 pp) |
| RoboCasa Composite Seen | RLDX-1 17.4%, π0.5 6.8% | 16.4% | 24.2% (+7.8 pp) |
| RoboCasa Composite Unseen | RLDX-1 6.8%, π0.5 0.9% | 4.7% | 13.3% (+8.6 pp) |
| RoboTwin, 50 tasks | HarnessVLA 58.0%, LingBot-VLA 50.4% | 48.5% | 74.0% |
Spatial, Object, and Goal rise in the early rounds. LIBERO-10 needs a longer edit sequence. The weakest cell is LIBERO-10 position-swap at 76.9%, where H0 is 0.0%. OpenVLA averages 0.0% and π0 averages 0.3% across the eight settings.
The active versus passive comparison uses four LIBERO-Goal tasks. The passive variant may read history, edit code, and run formal evals, but it cannot open a live scene during development. After one round, active search gains 50.0 percentage points at 42.2M meta-agent tokens per task; passive search gains 39.1 points at 94.2M tokens. After two rounds the gains are 85.9 versus 56.3 points, at 91.6M versus 137.7M cumulative tokens. The abstract says this also uses fewer simulator rollouts. The body quantifies tokens, not rollout steps.
Goal-t09 goes from 0% to 56.3% and then to 100%. The first round adds a lift check, re-segmentation, a bounded grasp retry, and an explicit transport-and-release. The remaining failures came from measuring lift against the approach pose. The fix measures displacement over the closed-contact segment, clears that evidence when contact breaks, and extends gripper opening from 8 control steps to 20. Goal-t08 shows the search is not monotonic: placement edits move the formal score from 50% to 68.75% and are promoted; a new stopping rule ties that score and is rejected; a side branch with a metric placement skill reaches 100% and scores 100% again on a frozen retest.
The editable surface is orchestration code. With π0.5 held fixed, a thin harness averages 17.5% on LIBERO-Pro and the evolved one averages 95.0%, above the 61.1% reported for HarnessVLA. The retained changes are checks, bounded recovery, and a few programmatic motion segments beside the VLA.
The bill is per task. Two rounds on four Goal tasks cost about 91.6M meta-agent tokens per task, with a frontier coding model in the loop. Composite RoboCasa tasks only reach 24.2% and 13.3%. On long household sequences this is a modest gain.
Treat it as paid, per-task harness development inside a simulator. The paper does not show that an evolved harness transfers to a new task.
Every number comes from high-fidelity simulation. No physical robot run is reported. The paper notes that real hardware adds sensor noise, actuation latency, and unmodeled contact, and it names Franka arms and mobile manipulators as follow-up, without results.
The development matrix is the scoring matrix. Formal episodes are fresh and unattended, but the task initializations match development, and each task receives its own harness. That measures how far code can be pushed on the evaluation set. Cross-task reuse is not reported.
The jump from 17.5% to 95% is against a thin initial interface. Against a mature harness, the relevant LIBERO-Pro figure is HarnessVLA at 61.1%. VISTA is still higher, and the paper does not show that the planner and tool stack match. On RoboCasa, H0 is already near RLDX-1, and evolution adds about 8 points, closer to another editing pass on an existing harness. On RoboTwin, H0 at 48.5% sits below the cited HarnessVLA result of 58.0%; 74.0% is what clears that baseline.
Active versus passive covers four tasks and two rounds, with no ablation that removes the version graph. A task's formal matrix is 8 seeds by 2 repetitions, so 68.75% is 11 successes out of 16. Promotion is a strict comparison of empirical means, with no significance bar.