Tsinghua and NUS Open-Source VISTA: Robots Self-Edit Code, Lifting Success Rate from 17.5% to 95%
A Tsinghua and National University of Singapore team proposed and open-sourced VISTA, a framework in which a Meta-agent iteratively improves an embodied agent's harness—skill code, invocation logic and persistent memory—through active exploration and code development, without updating model weights. On LIBERO-Pro it raises overall task success from 17.5% to 95.0%, with improvements retained and revisitable across rounds via a harness graph.
Full Report
Contents9 sections
VISTA: Driving Embodied Agents to Explore Actively and Iterate Their Capabilities
Contributed by the VISTA Team
QbitAI | WeChat Official Account QbitAI
Place a box of cheese on a shelf — the goal is clear, yet the robot must chain together localization, grasping, transport, and release.
How to tell that an object has been grasped securely, when to begin transporting it, and how to continue after a first failed attempt — all of this must be turned into concrete skills and decision logic.
So can this part of the work be handed over to the robot itself, letting it run its own experiments, modify its own code, and gradually accumulate experience to complete the task?
A research team from Tsinghua University and the National University of Singapore has proposed and open-sourced VISTA. It hands the work of experimentation and development to a Meta-agent, which acquires new evidence through active exploration and improves the Harness.
The entire pipeline requires no weight updates. With the underlying model parameters held fixed, VISTA raises the overall task success rate on LIBERO-Pro from 17.5% to 95.0%.
Treating the Harness as the Object of Evolution
An embodied Agent can use learned policies, motion planning, feedback control, and perception tools. The Harness is responsible for organizing these capabilities: what instructions to give the model, which skills to expose, how to invoke tools, and how to process execution feedback. Skill code, invocation logic, and persistent memory are all things that can be developed.
In VISTA, the execution Agent uses the Harness to complete tasks; the Meta-agent, acting as the developer, inspects the code and run records, writes or modifies skills, and then validates candidate versions. For example, it can combine localization, grasping performed by a learned policy, and transport executed by a controller into a new skill for the execution Agent to invoke. Improvements can therefore persist across rounds.
△ VISTA overview. The Meta-agent improves the Harness through environment interaction and code development.
How the Meta-agent Develops the Harness
Each development round starts from an existing Harness version and its execution records. The Meta-agent inspects the code, compares observations, actions, and task outcomes, identifies behaviors that need improvement, and proposes concrete modification ideas.
Around these issues, it modifies skills, instructions, or decision logic. If the existing records are insufficient for a judgment, it enters the simulation environment to observe and act, gathering supplementary evidence. During development it can also run autonomous tests and continue adjusting the candidate version based on execution results.
After submission, the candidate Harness is frozen and subjected to a new autonomous execution evaluation under preset task conditions. The Meta-agent does not participate in execution during evaluation; the candidate replaces the current version only if its results are better. Approaches tried during development are thus tested against actual task performance.
Code versions, inheritance relationships, and the corresponding evidence are stored together in the Harness graph. The next round can continue refining an effective version, or return to an earlier node to try other directions and reuse previously developed mechanisms. The skills and experience gained from one modification can therefore become the foundation for subsequent development.
How a Single Harness Improves Round by Round
Take one task as an example: the robot is required to place a cream cheese box on a shelf. The initial Harness failed all 16 development evaluations on this task.
In the first development round, the Meta-agent built an explicit transport-and-release procedure for this task: first confirm that the object has been lifted, then perform transport; if grasping fails, re-localize the object and retry within a limited number of attempts. These modifications were written into the Harness, and the submitted version's success count rose to 9/16.
The second round retained the already effective procedure and continued analyzing the remaining failures. The execution records exposed a specific problem: the starting point for computing lift displacement was inconsistent with the actual grasp-contact phase, which affected the judgment of grasp state. The revised code computes displacement along the continuous closed-contact phase, resets the evidence when contact is lost, and also adjusts the release behavior. The new version reached 16/16, and when the same version was evaluated again, it also achieved 16/16.
△ The skill development process.
This process left behind skills that could be further developed: transport, release, and limited retries were retained first, and later versions then refined the state judgments within them. Each development round has concrete code and execution results as the starting point for the next round.
Experimental Results of Harness Evolution
The experiments cover LIBERO, RoboCasa, and RoboTwin. The table below compares the success rates of the initial Harness and the evolved version selected for each task. During evaluation, the Harness code is fixed, and the execution Agent completes tasks autonomously.
The average success rates on LIBERO-T and LIBERO-S rise from 22.8% and 12.2% to 96.6% and 93.4%, respectively; RoboTwin's 50 tasks rise from 48.5% to 74.0%. In RoboCasa's multi-stage tasks, the success rates on Seen and Unseen rise from 16.4% and 4.7% to 24.2% and 13.3%, respectively.
The round-by-round results break the overall improvement down to individual tasks. Among the ten tasks in Goal-T, eight already produced 16/16 evaluation results within the first three development rounds. The development pace still differs markedly across tasks, and the performance of later candidates may also fluctuate, so evaluation is needed to decide which version to keep.
△ Round-by-round results on LIBERO Goal-T.
Toward Recursive Self-Evolution of Embodied Agents
Once model training is complete, the robot's capabilities need not be fixed from then on. VISTA lets a Coding Agent continuously develop the Harness, turning experience gained during execution into skills that subsequent versions can inherit. Current simulation experiments show that this kind of development can deliver real improvements while model parameters remain unchanged. Going further, we hope robots in real environments can also gradually take on this part of the work: developing, validating, and refining their own skills based on task experience, so that capability growth continues after deployment. Paper link: https://arxiv.org/abs/2610.35051 Code repository: https://github.com/AgenticEI/VISTA*
One-click triple action: "Like," "Share," "Heart"
Feel free to leave your thoughts in the comments!
— The End —
【Academic Submissions】Please send an email on weekdays to: [email protected], with 【Submission】 in the subject line, and tell us: who you are, where you are from, attach a project/homepage link to the submission content, and provide contact information.
🎓 We will (try our best to) reply to you promptly :)
🌟 Turn on the star 🌟
Daily frontier tech updates