Embodied agents are increasingly used to close the gap left by end-to-end policy models. Yet the agentic path has not realized closed-loop learning in physical execution: existing harnesses remain largely open-loop, following fixed skills during rollout and reflecting only after an episode completes. Such post-hoc reflection cannot govern execution as it unfolds, because physical interaction requires decisions to track rapidly changing robot-environment states at a frequency beyond today's large agentic models. We present Zetta, a closed-loop embodied harness that evolves code-based runtime critics and recovery skills online while keeping the base policy frozen. Through three timescale-separated loops, Zetta provides action-frequency governance, rollout-level critic-recovery proposal, and validation-gated skill updates. Together with Z-Infra, a rollout infrastructure decoupling agent logic from heterogeneous execution resources, Zetta achieves state-of-the-art success on LIBERO-Pro and RoboCasa under our current rollout budget, reaching 90.8% and 93.6%, with an 11.1x inference speedup; success continues to scale with self-exploration experience; learned skills transfer zero-shot, and clear robotic "Aha Moments" emerge. These results show that closed-loop harness self-evolution opens a scaling path for reliable physical intelligence.
Alex: Welcome to another episode of ResearchPod. Today we're looking at a paper introducing Zetta, a system designed to help robots learn from their own physical mistakes over time.
Sam: So the core problem is that robots keep making the same errors when working in messy, real-world environments — like a kitchen — and can't seem to correct themselves?
Alex: Exactly. The underlying issue is that modern robot brains are surprisingly good at understanding language and high-level goals, but once they start moving, they typically operate in what researchers call an "open-loop" way. Think of it like following a recipe with your eyes closed. You've planned every step in advance, but if something shifts on the counter, you have no way of noticing or adjusting.
Sam: So if a cup slides a few centimetres to the left, the robot just keeps reaching for where it thought the cup was?
Alex: Right. And that's the cascade problem. A small slip early on — something a human would correct without thinking — compounds into a complete task failure because the system has no mechanism to notice things have gone wrong and adapt.
Sam: I suppose the obvious fix would be to retrain the robot's brain every time it fails. But that can't be practical.
Alex: It isn't. These foundation models — the large neural networks that give robots their language understanding and general reasoning — take enormous amounts of computing power and time to train. You can't restart that process every time a gripper misses a cup. So instead of modifying the core model, Zetta wraps a separate, lightweight layer around it. The foundation model stays frozen. The new layer handles the monitoring and recovery.
Sam: That's a useful distinction. The brain stays the same — you're just adding a more attentive nervous system around it.
Alex: That's a good way to put it. And that outer layer works by building up a library of recovery behaviours over time. When the robot encounters a failure, the system doesn't just log it and move on. It tries to understand what went wrong and constructs a reusable skill for handling that situation in the future.
Sam: So it's less like the robot is getting smarter in the moment, and more like it's building up a catalogue of lessons it can draw on later?
Alex: Precisely. And the way they validate whether those lessons actually stuck is quite rigorous. The system has to pass two separate tests before any update is accepted. First, it must successfully replay every single scenario it previously failed — with a perfect success rate, not just an improvement. Second, it has to perform well on tasks it has never seen before, to prove it has genuinely learned something general rather than just memorising a specific fix.
Sam: That second test matters a lot. A robot that only gets better at the exact mistakes it's already made isn't really learning — it's just patching.
Alex: Exactly. And if the evaluation reveals new failure modes that weren't there before, the system flags them and demands fresh validation before anything is deployed. It's a conservative, careful cycle by design.
Sam: What about the computing side of things? Running all these simulations in parallel sounds like it would put serious strain on the hardware.
Alex: It does, and that's why the paper introduces a separate piece of infrastructure called Z-Infra. The challenge is that robotic agents don't use computing resources in a smooth, predictable way. They make decisions dynamically — choosing which tools to use, when to act, when to pause — and that creates what the authors describe as bursty demand on the hardware. One moment the system is idle, the next it needs a surge of processing power.
Sam: So you can't just allocate a fixed amount of computing and call it done.
Alex: Right. Z-Infra handles this by separating the agent's decision-making logic from the hardware management entirely. There's a coordinating layer that routes requests to specialised workers — some handling the simulation environment, others handling the model's inference. And crucially, the vision and language components of the model run as separate processes that communicate through fast memory transfers, which keeps response times low.
Sam: So the system is designed to be modular — each piece doing one job well, rather than one monolithic process trying to do everything.
Alex: That's the design philosophy, yes. And it matters because it makes the whole system easier to scale. You can add more workers, swap out components, or adjust resource allocation without rebuilding everything from scratch.
Sam: Stepping back — what does this paper actually claim to demonstrate? Is this tested on real robots, or is it still in simulation?
Alex: The paper is candid about this. The current results come from simulation, and the authors identify bringing these findings to physical hardware as the clear next step. So the claims are about the framework's design and its performance in controlled conditions, not yet about deployment in the real world.
Sam: That's an important caveat. Simulation and physical reality can diverge in ways that are hard to predict.
Alex: They can. The gap between a simulated kitchen and an actual one — with variable lighting, imperfect surfaces, and objects that don't behave exactly as modelled — is a genuine challenge the field is still working through. What this paper contributes is a structured approach to the recovery and validation problem, which is a meaningful piece of that larger puzzle.
Sam: So the value isn't a finished solution, but a more principled way of thinking about how robots should handle failure.
Alex: That's a fair summary. Rather than treating failure as a reason to retrain everything from the ground up, Zetta treats it as information — something to be catalogued, understood, and turned into a reusable skill. Whether that approach holds up on physical hardware is the open question, but the framework itself offers a more practical path than the alternatives.
Sam: Thanks for walking us through it. Thanks for listening to ResearchPod.