cs.ROAug 17, 2026
SaveZetta : An Efficient Closed-Loop Embodied Harness for Self-Evolving Physical Intelligence
Organizations: Institute for AI Industry Research (AIR), Tsinghua University; Z-Trans AI
Abstract
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.
Figures & tables
Figure 1: Zetta closes the loop for embodied self-evolution. Frequent runtime critics trigger recoveries during execution, while verified failures are distilled into reusable critic and recovery skills across rollouts. A hardware-decoupled rollout layer scales this process across heterogeneous environments, models, CPUs, and GPUs. Zetta enables sustained same-task improvement, zero-shot skill transfer, robotic "Aha moment", and accelerated agent execution.
Figure 2: Overview of the Zetta evolutionary framework. The system operates in a continuous loop between online execution and offline evolution. (Left) Parallel Rollouts: Action policy executes tasks, monitored by an Evolvable Harness ( ) composed of Critics ( ), Recovery ( ), and Tools ( ), all under the high-level adjudication of an Orchestrator Agent . Rollouts are categorized into success and failure trajectories. (Right) Reflection & Evolve: Failed trajectories trigger a three-phase offline evolution cycle. Phase I profiles failures by clustering them against successful reference baselines. Phase II performs causal diagnosis to locate the root failure layer ( ), followed by minimal harness repair to update , , and . Phase III generalizes seed-specific repairs into a unified, versioned Harness Package ( ), which is then fed back into the execution loop.
Figure 3: Three-layer architecture of the rollout infrastructure. The Control Plane routes agent requests to specialized Env Workers and Rollout Workers, which manage environment simulation and model inference respectively.
| API Primitive | Functionality |
| Session Management — create once, run many episodes, close when done | |
| create_sessions(requests) | Batch create sessions with env_family, env_config, lease_seconds. Returns SessionHandle with session_id. |
| renew_sessions(session_ids) | Extend lease duration for active sessions. |
| close_sessions(session_ids) | Close sessions and release resources (idempotent). |
| Environment Control — direct interaction for custom control logic | |
| reset(session_ids, reset_spec) | Start new episode with task_id, seed, instruction. Returns episode_id and initial observation. |
Table 1: Core API primitives exposed to agents. All primitives support batching across sessions.
Figure 4: Physical-intelligence “Aha” moments on LIBERO-Pro. Similar to RoboCasa, v0 denotes the Pure-VLA baseline and v1 represents early symptomatic repairs (e.g., staging or local gates) that yield stagnant performance. The “Aha” at v2 occurs when the agent identifies and resolves the decisive physical bottleneck (e.g., grasp retention or semantic approach), triggering a sharp increase in execution reliability.
Figure 5: Physical-intelligence “Aha” moments on RoboCasa. L2-v0 is the original Pure-VLA baseline, while L2-v1 is an intermediate internal version. The flat L2-v0–L2-v1 segments summarize a period where early candidate fixes overfit individual failures. Once the agent identifies the key physical bottleneck (e.g., EEF alignment or centered contact), L2-v2 produces a sharp success-rate increase.
Figure 6: Cumulative physical-intelligence scaling on LIBERO-Pro Goal. Average performance across ten tasks under Goal-T and Goal-S perturbations. Each point denotes a selected cumulative harness version. The base VLA is frozen throughout evolution.
Figure 7: Cross-task scaling from the Goal-T8 source task. Three cumulative Critic–Recovery capabilities discovered on Goal-T8 are transferred to Goal-T2, Goal-T6, and Goal-S3. Curves report complete fixed-seed evaluations of each cumulative stack. Shaded regions denote Wilson confidence intervals. The right-hand frames illustrate the corresponding failure and recovery mechanisms rather than additional transfer measurements.
Figure 8: Cross-task scaling from the Goal-S5 source task. Cumulative Critic–Recovery capabilities discovered on Goal-S5 are transferred to Goal-S3, Goal-S4, and Goal-S9. Within each task, all four arms use the same fixed seeds, policy RNGs, checkpoint, and execution budget. Shaded regions denote Wilson confidence intervals.
Figure 9: Cumulative reflection scaling across 18 RoboCasa tasks. The macro-average success rate increases from 73.56% for the frozen parent harness to 78.71%, 84.85%, 90.54%, and 93.56% after four cumulative global repair rounds. Each checkpoint retains the previously validated critic, recovery, and tool capabilities.
Figure 10: Reflection-driven scaling and transfer on PnP tasks. On PnP-Stove, successive rounds add object-relative pregrasp alignment, bounded regrasp after grasp loss, and stable placement. The lower panels apply the cumulative checkpoints to PnP-Sink, PnP-Cabinet, and PnP-Toaster without an additional training or evolution loop. The right-hand panels show the failure signatures identified by the critic and the corresponding recoveries.
Figure 11: Reflection-driven scaling and transfer on articulated interaction tasks. On TurnOffStove, successive rounds add target localization, collision-aware pre-contact approach, and stable EEF–target contact. The lower panels apply the cumulative checkpoints to TurnOnSinkFaucet, OpenCabinet, and TurnOnMicrowave without an additional training or evolution loop. The right-hand panels show the diagnosed target, approach, and contact failures together with the recovered states.
| Method | T1 | T2 | T3 | T4 | T5 | T6 | T7 | T8 | T9 | Avg. |
| Pure VLA (GR00T) | 78 | 74 | 78 | 58 | 48 | 62 | 72 | 74 | 70 | 73.56 |
| Zetta | 96 | 92 | 94 | 96 | 86 | 80 | 96 | 96 | 86 | 93.56 |
| Method | T10 | T11 | T12 | T13 | T14 | T15 | T16 | T17 | T18 | Avg. |
| Pure VLA (GR00T) | 62 | 92 | 76 | 88 | 96 | 50 | 90 | 82 | 74 | 73.56 |
| Zetta | 94 | 98 | 94 | 94 | 100 | 100 | 94 | 96 | 92 | 93.56 |
Table 2: Success rates (%) on 18 RoboCasa Atomic-Seen tasks. Task identifiers follow Appendix A . The best result for each task is shown in bold. “Avg.” is the macro-average over all 18 tasks.
| Setting | Method | Task 0 | Task 1 | Task 2 | Task 3 | Task 4 | Task 5 | Task 6 | Task 7 | Task 8 | Task 9 | Average |
| Goal (T) | 0.0 | 95.0 | 10.0 | 0.0 | 100.0 | 0.0 | 20.0 | 80.0 | 5.0 | 0.0 | 31.0 | |
| Zetta | 80.0 | 100.0 | 95.0 | 80.0 | 100.0 | 100.0 | 95.0 | 95.0 | 80.0 | 100.0 | 92.5 | |
| Goal (S) | 0.0 | 60.0 | 0.0 | 45.0 | 0.0 | 0.0 | 0.0 | 100.0 | 100.0 | 75.0 | 38.0 | |
| Zetta | 90.0 | 65.0 | 80.0 | 85.0 | 95.0 | 95.0 | 100.0 | 100.0 | 100.0 | 80.0 | 89.0 | |
| LIBERO-10 (T) | 5.0 | 95.0 | 95.0 | 0.0 | 0.0 | 80.0 | 85.0 | 75.0 | 65.0 | 0.0 | 50.0 | |
| Zetta | 35.0 | 95.0 | 100.0 | 0.0 | 25.0 | 100.0 | 95.0 | 80.0 | 100.0 | 0.0 | 63.0 |
Table 3: Success rates (%) on LIBERO-Pro. The best result for each task is shown in bold. “Average” is the macro-average over 10 tasks.
Figure 12: A sequence of critic–recovery interventions during aPnP episode. The VLA first picks and transports the object. A critic detects a lost grasp and triggers a re-approach recovery; an invalid re-grasp pose then triggers GraspGen to synthesize a feasible grasp pose; near the goal, a final critic invokes stable CAP placement. After each local repair, control returns to the nominal VLA only after the recovery state is verified.
Figure 13: Critic–recovery promotion progressively moves the failure frontier in LIBERO-Pro Goal-S5. The parent VLA exhausts its execution budget without a timely recovery handoff. Successive promotion rounds introduce a retention-gated handoff, a contact-gated grasp recovery, and a bounded re-grasp retry. Each addition resolves the previously exposed failure and reveals a later-stage bottleneck, until retained transport resumes and the official stove-front relation is satisfied.
Figure 14: Per-episode latency versus concurrency. Z-Infra maintains controlled growth while baselines exhibit explosive degradation or inherently high overhead. Both baselines encounter OOM beyond concurrency 16.
Figure 15: Rollout throughput versus concurrency. Z-Infra scales to 35.1 ep/min at concurrency 64, achieving 7.7× and 12.8× higher throughput than baselines at concurrency 16. Baselines fail beyond concurrency 16 due to OOM.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
| ID | Official task name |
| T1 | NavigateKitchen |
| T2 | TurnOnMicrowave |
| T3 | PickPlaceCounterToStove |
| T4 | PickPlaceSinkToCounter |
| T5 | PickPlaceDrawerToCounter |
| T6 | PickPlaceCounterToCabinet |
Table 4: Mapping from compact identifiers to official RoboCasa task names.
| ID | LIBERO-Goal | ID | LIBERO-10 (Long) |
| Task 0 | open the middle drawer of the cabinet | Task 0 | put both the alphabet soup and the tomato sauce in the basket |
| Task 1 | put the bowl on the stove | Task 1 | put both the cream cheese box and the butter in the basket |
| Task 2 | put the wine bottle on top of the cabinet | Task 2 | turn on the stove and put the moka pot on it |
| Task 3 | open the top drawer and put the bowl inside | Task 3 | put the black bowl in the bottom drawer of the cabinet and close it |
| Task 4 | put the bowl on top of the cabinet | Task 4 | put the white mug on the left plate and put the yellow and white mug on the right plate |
| Task 5 | push the plate to the front of the stove | Task 5 | pick up the book and place it in the back compartment of the caddy |
Table 5: Mapping of LIBERO-Pro task identifiers used in the main text to the underlying LIBERO tasks. “Goal” corresponds to LIBERO-Goal and “Long” corresponds to LIBERO-10. The same task indices are used for both T and S perturbation settings.