Zetta ζ: An Efficient Closed-Loop Embodied Harness for Self-Evolving Physical Intelligence

Summary (Overview)

• Zetta introduces a closed-loop embodied agent harness that evolves code-based runtime critics and recovery skills online while keeping the base policy model (VLA/WAM) frozen, enabling sustained same-task improvement without additional policy training.
• The system achieves state-of-the-art success rates of 90.8% on LIBERO-Pro and 93.6% on RoboCasa under the current rollout budget, with an 11.1× inference speedup compared to the RPent baseline.
• Key contributions include: (1) the first self-evolving embodied agent through learnable code-based critics and three coordinated evolution loops; (2) the first rollout infrastructure (Z-Infra) designed specifically for embodied self-evolution; (3) demonstrated zero-shot skill transfer across tasks and clear robotic "Aha Moments."
• Performance scaling: Success increases from 34.5% → 90.8% on LIBERO-Pro and from 73.6% → 93.6% on RoboCasa within several evolution iterations.
• Zero-shot transfer: Skills discovered on source tasks (e.g., PnP-Stove) improve success on related tasks (PnP-Sink: 58% → 82%, PnP-Cabinet: 62% → 80%, PnP-Toaster: 72% → 90%) without any target-specific optimization.

Introduction and Theoretical Foundation

Background and Motivation

Physical intelligence is advancing along two complementary paths:

1. End-to-end policy models (VLAs, WAMs) trained on large demonstration corpora—but these are brittle under real-world distribution shifts.
2. LLM-based embodied agents that orchestrate policy models, code, tools, and control primitives—but these remain open-loop during execution.

The fundamental bottleneck is that physical tasks require decisions at millisecond-level frequency, which large agentic models cannot achieve. Existing systems perform reflection only at episode level, which:

• Cannot govern execution as it unfolds
• Cannot test alternative actions to verify reflections
• Makes credit assignment over entire trajectories difficult
• Lacks access to precise failure states

Theoretical Foundation: Dual-Agent Governance

The framework formalizes long-horizon robotic manipulation as a governance and evolution process with:

Fixed Runtime Components:

• Action Policy (π): Generates low-level actions $a_t = \pi(s_t, g; \theta)$ with constraint $\nabla\theta = 0$
• Orchestrator Agent ($\mathcal{A}_{orch}$): Fixed multimodal reasoning operator for adjudication

Evolvable Harness (H):

$$\mathcal{H} = \{C, R, \mathcal{T}\} \tag{1}$$

• Runtime Critic (C): High-frequency monitoring functions that generate structured proposals:

$$P_t = C(\tau_{0:t}) = \langle e_t, \hat{\sigma}_t \rangle \tag{2}$$

• Recovery Playbook (R): Structured library of strategies mapped to failure mechanisms
• Heterogeneous Toolset (T): Executable tools and operators that can be generated and refined

Adjudication Logic (online execution):

$$\sigma_t = \mathcal{A}_{orch}(P_t, R, \mathcal{T}, \mathcal{K}) \tag{3}$$

Offline Optimization (evolutionary agents):

$$\mathcal{H}^{(k+1)} \leftarrow \mathcal{A}_{evo}(\mathcal{D}_{fail}^{(k)}, \mathcal{H}^{(k)}) \tag{4}$$

Optimization Objective:

$$\max_{\mathcal{H}} J(\mathcal{H}) = \mathbb{E}_{g, s_0 \sim \mathcal{D}} [\text{Success}(\tau) \mid \pi, \mathcal{A}_{\text{orch}}, \mathcal{H}] \tag{5}$$

Three Core Challenges

1. Open-Loop Semantics-Physics Gap: VLA/WAM models lack closed-loop sensory-motor perception → Solution: High-frequency state governance via critics
2. Generalization Pitfall of Over-Parameterized Repair: Overfitting repairs corrupt action distributions → Solution: Top-down hierarchical causal diagnosis with minimal intervention
3. Scalability of Expert-in-the-Loop Debugging: Manual debugging is unscalable → Solution: Fully automated evolutionary learning loop

Methodology

Three Coordinated Evolution Loops

Loop 1: Critic-Governed Action Loop — Executes learned critics at action frequency, triggers appropriate recovery skills when necessary

Loop 2: Rollout-Batch Candidate Optimization Loop — Clusters and diagnoses failures, proposes candidate critics and recoveries using an SGD-like optimization in code space (drawing on SkillOpt and EmbodiSkill)

Loop 3: Validation-Gated Skill Update Loop — Admits only candidates that improve success rate and generalize across rollouts into skill memory

Phase I: Empirical Failure Profiling

• Large-scale sampling of the action policy across development seed set $\mathcal{D}_{dev}$
• Deterministic Scheduling: Centralized resource scheduler routes tasks with identical containers/hardware
• Validity Determination: Valid rollout set $\mathcal{V}$ includes only complete trajectories with full sensor/video evidence
• Multi-Dimensional Evidence: Each trajectory is a multimodal time series:

$$\boldsymbol{\tau}^{(j)} = \{(s_t, a_t, \mu_t, \phi_t)\}_{t=0}^T \tag{7}$$

• $s_t$: state, $a_t$: action, $\mu_t$: milestone completion status, $\phi_t$: physio-auxiliary signals

Phase II Stage 1: Failure Clustering & Causal Diagnosis

First Missing Milestone ($m^*$): The first milestone not observed in trajectory history:

$$m^* = \min\{m_k \in M \mid m_k \notin \{\mu_t\}_{t=0}^T\} \tag{9}$$

Earliest Observable Divergence ($t_{EOD}$): First time where state deviates from healthy distribution:

$$t_{EOD} = \min\{t \mid \text{dist}(s_t, s_t^{ref}) > \epsilon, s_t^{ref} \in I_{succ}(\mu_t)\} \tag{10}$$

Top-Down Hierarchical Diagnosis: Traverses layers in priority order: Evaluation → Critic → State → Planning/Control → Recovery → Parameter

Phase II Stage 2: Harness Repair & Validation

VLA Re-entry Contract: Defines when control returns to base policy:

$$\Psi(s_t) = \mathbb{1}(\text{FailureCleared}) \wedge \mathbb{1}(\text{Stability}(s_t) > \gamma) \tag{12}$$

Validation Success Criteria:

$$\text{Success}(\mathcal{H}_{\text{patch}}) = \mathbb{1}(\mu_{T,\text{new}} = m_{\text{goal}} \wedge \forall t \in \text{Intv}, \text{Adjudicated by} \mathcal{A}_{\text{orch}}) \tag{13}$$

Phase III: Harness Consolidation & Generalization

• Historical Regression: All failed seeds in cluster must succeed (100% success rate)
• Held-out Evaluation: Success rate increment over baseline:

$$\Delta SR = SR(\mathcal{D}_{\text{held-out}} \mid \mathcal{H}_{\text{merged}}) - SR(\mathcal{D}_{\text{held-out}} \mid \pi_{VLA}) \tag{16}$$

• Dynamic Transition: New failure mechanisms trigger reclassification of held-out seeds as development seeds

Z-Infra: Three-Layer Rollout Infrastructure

1. Control Plane: Gateway with unified API, session registry, heartbeat monitoring, fault tolerance
2. Environment Worker: Manages simulation environments with session-based lifecycle, resource-sharing groups (ModelTemplate + fork), high-performance C++ controller
3. Rollout Worker: GPU-resident model serving with batched inference, dynamic scheduling, model partitioning (VLM/AE separate processes, 53% latency reduction), quantization runtime (W8A8 for prefix MLP, 1.18×–1.32× speedup)

Empirical Validation / Results

"Aha" Moments (Discontinuous Capability Gains)

Scaling Results

LIBERO-Pro: Average success increases from 31.0% → 92.5% (Goal-T) and 38.0% → 89.0% (Goal-S) through cumulative Critic–Recovery mechanisms.

RoboCasa: Macro-average success increases from 73.56% → 93.56% across four cumulative global repair rounds.

Final Benchmark Results

RoboCasa Atomic-Seen Tasks (18 tasks):

LIBERO-Pro (40 task-setting pairs):

Z-Infra Performance

• Throughput: 35.1 episodes/min at concurrency 64 (vs. 2.88 ep/min for without Z-Infra, 1.72 ep/min for RPent)
• Latency: 95s per episode at concurrency 64 vs. 513s for RPent (11.9× reduction)
• Inference: 11.1× speedup, 53% latency reduction from model partitioning, 1.18×–1.32× from quantization

Theoretical and Practical Implications

Theoretical Significance

1. Closed-loop governance paradigm: The work demonstrates that frozen VLA/WAM policies can be made reliable through evolvable code-based critics, challenging the assumption that policy improvement requires weight updates.

2. Scaling unit insight: The "scalable unit of physical intelligence" is identified as state variables (grasp stability, approach geometry) rather than task-specific trajectories — offering a new theoretical framework for understanding what constitutes reusable knowledge in embodied AI.

3. Three-timescale evolution: The separation of action-level (milliseconds), rollout-batch-level (seconds/minutes), and iteration-level (hours) loops provides a principled decomposition for self-evolving agent systems.

Practical Implications

1. Deployment-time self-improvement: Harness self-evolution enables robots to improve through experience in real-world deployment without requiring model retraining, which is significant for applications where data collection is expensive or safety-critical.

2. Infrastructure as enabler: Z-Infra demonstrates that rollout throughput is the rate-limiting factor for intelligence scaling — faster rollouts directly accelerate the evolution loop, making infrastructure design a primary research problem.

3. Zero-shot transfer: Skills learned from a single task can be transferred to related tasks, suggesting that harness-based skill acquisition could replace task-specific training for common manipulation primitives.

Conclusion

Main Takeaways

Zetta establishes a new paradigm for physical intelligence scaling: instead of scaling end-to-end policy models, it scales the governance harness around frozen policies. By evolving code-based runtime critics and recovery skills through three coordinated loops, the system achieves:

• Sustained same-task improvement with frozen policy weights
• Zero-shot capability transfer across tasks
• Discontinuous "Aha" moments when the true physical bottleneck is identified
• Accelerated execution through efficient rollout infrastructure

Future Directions

The paper identifies clear next steps:

1. Real robot deployment: Extending Zetta and Z-Infra to real robots by bridging the sim-to-real gap
2. First-class real-robot environments: Integrating real-robot environments as first-class workers alongside simulated ones in Z-Infra
3. Massively parallel rollout collection: Enabling fast, parallel rollout collection for self-evolution directly with real machines

The work establishes that harness self-evolution is a viable path toward reliable embodied intelligence, complementing the end-to-end policy scaling approach with a more sample-efficient, safer alternative that leverages existing frozen models while continuously improving execution reliability.