Why real-time correction matters for long tasks
The main reason closed-loop harnesses can avoid repeating mistakes is that they correct errors while the task is still unfolding, not just after it finishes. Zetta, a closed-loop embodied harness, uses three timescale-separated loops to govern actions at high frequency, propose critic-recovery skills during rollout, and validate skill updates before adopting them [1]. This means the system can catch a wrong move immediately and adjust, rather than waiting until the end of the episode to reflect. In tests, Zetta achieved 90.8% success on LIBERO-Pro and 93.6% on RoboCasa, benchmarks that involve long, multi-step manipulation tasks [1]. These numbers show that real-time correction dramatically improves reliability compared to open-loop systems that only reflect post-hoc.
The role of physics and causality in planning
Another critical piece is giving the harness an innate understanding of physics and causality, so it can simulate the consequences of its actions before committing to them. The closed-loop multi-step planning framework does exactly this: it defines temporary closed-loop controllers called 'Tasks' and a supervisory 'Configurator' that uses a physics engine to simulate sequences of tasks and store results in a model of the environment [2][3]. This allows the robot to plan by chaining tasks that are contingent on sensory input, and to avoid mistakes by predicting what will happen next. The framework was implemented on a real robot and tested in an overtaking scenario, proving that this approach works in practice [3]. Together with Zetta's real-time correction, these two approaches—reactive correction and predictive simulation—complement each other and together address the challenge of long-horizon reliability.
What the evidence doesn't yet tell us
While the results are promising, there are important caveats. Zetta's success was achieved under a specific rollout budget and with a frozen base policy, meaning the harness evolves its critics and recovery skills but not the underlying policy [1]. This raises questions about whether the approach scales to even more complex tasks or to tasks where the base policy itself is flawed. The closed-loop multi-step planning framework was tested in only two scenarios (overtaking and another proof-of-concept), so its generalizability is still unknown [2][3]. Also, neither paper reports long-term retention of learned skills or performance on tasks with hundreds of steps, so we can't yet say whether these systems avoid mistakes indefinitely. The evidence strongly suggests closed-loop harnesses are a step forward, but the field is still early, and more testing on diverse, longer tasks is needed.
About These Sources
This answer is built on 3 studies (all preprints) — published from 2024 to 2026, 3 from 2024 or later — selected as the most relevant from 3 studies that passed quality screening, drawn from 51 papers retrieved from a database of over 500 million.
Sources used in this answer
Zetta ζ : An Efficient Closed-Loop Embodied Harness for Self-Evolving Physical Intelligence
Zetta, a closed-loop embodied harness, achieved 90.8% success on LIBERO-Pro and 93.6% on RoboCasa by evolving runtime critics and recovery skills online while keeping the base policy frozen, with an 11.1x inference speedup and zero-shot transfer of learned skills.
Closed-loop Multi-step Planning
The closed-loop multi-step planning framework defines temporary closed-loop controllers ('Tasks') and a supervisory module with innate physics and causality knowledge, enabling planning by chaining tasks; it was implemented on a real robot and tested in two scenarios as proof of concept.
Closed-loop multi-step planning with innate physics knowledge
The hierarchical framework with a physics-engine-based Configurator was implemented on a real robot and tested in an overtaking scenario, demonstrating that closed-loop multi-step planning with innate physics knowledge works in practice.