Learning Without Losing Identity: Capability Evolution for Embodied Agents

Abstract: Embodied agents are expected to operate persistently in dynamic physical environments, continuously acquiring new capabilities over time. Existing approaches to improving agent performance often rely on modifying the agent itself -- through prompt engineering, policy updates, or structural redesign -- leading to instability and loss of identity in long-lived systems. In this work, we propose a capability-centric evolution paradigm for embodied agents. We argue that a robot should maintain a persistent agent as its cognitive identity, while enabling continuous improvement through the evolution of its capabilities. Specifically, we introduce the concept of Embodied Capability Modules (ECMs), which represent modular, versioned units of embodied functionality that can be learned, refined, and composed over time. We present a unified framework in which capability evolution is decoupled from agent identity. Capabilities evolve through a closed-loop process involving task execution, experience collection, model refinement, and module updating, while all executions are governed by a runtime layer that enforces safety and policy constraints. We demonstrate through simulated embodied tasks that capability evolution improves task success rates from 32.4% to 91.3% over 20 iterations, outperforming both agent-modification baselines and established skill-learning methods (SPiRL, SkiMo), while preserving zero policy drift and zero safety violations. Our results suggest that separating agent identity from capability evolution provides a scalable and safe foundation for long-term embodied intelligence.

• The paper introduces a novel capability evolution paradigm that decouples the persistent agent identity from evolving Embodied Capability Modules.
• The paper applies a hybrid learning methodology, combining RL, imitation, and LLM synthesis, to drive safe and stable evolution of skills.
• The paper demonstrates significant improvements with task success rates rising from 32.4% to 91.3%, while maintaining zero agent policy drift and enforcing strict safety constraints.

Capability Evolution Paradigm for Embodied Agents

Introduction

The paper "Learning Without Losing Identity: Capability Evolution for Embodied Agents" (2604.07799) formalizes a novel paradigm for constructing embodied agents that enables continuous skill acquisition and runtime safety, while maintaining strict invariance of agent identity. This approach explicitly decouples the agent’s cognitive core from its evolving behavioral competencies, operationalized through the introduction of Embodied Capability Modules (ECMs). The method is validated via extensive experiments on multi-step robotic manipulation tasks, demonstrating sustained performance improvement, system stability, and enforced safety constraints.

Formal Framework: Decoupling Agent Identity and Capability

The core abstraction is the separation between a persistent agent and a set of modular, versioned ECMs. The agent, characterized by a fixed policy $\pi_{\mathcal{A}}$ and identity memory $\mathcal{M}_{\mathcal{A}}^{\text{id}}$, is strictly immutable. All episodic learning traces are stored separately in $\mathcal{M}_{\mathcal{A}}^{\text{ep}}$, ensuring that ongoing experience accumulation does not result in drift of the agent’s essential cognitive parameters.

ECMs encapsulate all learnable and adaptable capacities: each module implements a specification $\phi_c$, parameterization $\theta^{(k)}$, and explicit input/output interfaces, and maintains a full version registry, allowing for lifecycle management (creation, deployment, iterative evolution, rollback, and deprecation). Importantly, agent-level reasoning only plans and orchestrates the invocation and composition of ECMs; it does not directly learn or adapt.

This decomposition ensures identity invariance: for all $t$, $$(\pi_{\mathcal{A}}^{(t)}, \mathcal{M}_{\mathcal{A}}^{\text{id}, (t)}) = (\pi_{\mathcal{A}}^{(0)}, \mathcal{M}_{\mathcal{A}}^{\text{id}, (0)})$$, while all change manifests solely in the ECM set $\mathcal{C}(t)$. The agent acts as a stable orchestrator, with all skill evolution and adaptation externalized.

Task execution is strictly gated through a runtime governance layer enforcing hard safety and policy constraints; this layer blocks, modifies, or logs every ECM invocation, and is architecturally independent from both the agent and capability modules.

Methodology: ECM Lifecycle, Learning, and Governance

ECMs follow a formal lifecycle:

• Creation: ECMs are instantiated via manual, LLM-based code synthesis, or cloning.
• Deployment: ECMs must satisfy interface compatibility and are added to the active capability set.
• Evolution: ECMs collect execution traces and are evolved using reinforcement learning (RL), imitation learning, LLM-based code updates, or hybrid combinations. New versions are deployed only if gated policy checks are passed.
• Deprecation: Superseded or obsolete ECMs are removed from the active set but retained for auditability and rollback.

Learning is performed in a closed loop, where performance and failures drive updates of ECMs, but the agent policy and identity parameters are never altered. This enforces the architectural guarantee of identity preservation and modular improvement.

Empirical Evaluation

The framework is validated on a challenging suite of multi-step robotic manipulation tasks in the robosuite/MuJoCo environment, with stochastic perturbations. Six tasks of varying complexity (pick, place, stack, pour, sort, assemble) are used to assess both single-ECM and compositional ECM capabilities.

Five approaches are compared:

• Static ECM: No skill evolution; agent observes but cannot adapt.
• Agent Modification: Agent-level prompt and policy adaptation, ECMs fixed.
• SPiRL: Pretrained skill priors, frozen after extraction; only agent adapts.
• SkiMo: Joint skill and world-model updates; agent and skills both mutable.
• Capability Evolution (this framework): Only ECMs evolve; agent is strictly immutable.

Key Numerical Results

• Capability evolution achieves a rise in task success rate from 32.4% to 91.3% over 20 iterations.
• The variance in performance ($\sigma^2$) is minimized (2.0) compared to all baselines, demonstrating system-level stability.
• Zero agent policy drift ($\Delta\pi=0$) and zero safety violations are enforced, distinguishing this approach from both agent modification methods (where $\mathcal{M}_{\mathcal{A}}^{\text{id}}$0 and $\mathcal{M}_{\mathcal{A}}^{\text{id}}$1 is substantial) and coupled architectures (e.g., SkiMo).
• Cohen's $\mathcal{M}_{\mathcal{A}}^{\text{id}}$2 effect sizes for the performance difference (Capability Evolution vs. Agent Modification and SkiMo) are in the range of 1.9-3.7, and all differences are statistically significant ($\mathcal{M}_{\mathcal{A}}^{\text{id}}$3).
• Runtime policy enforcement eliminates unsafe executions entirely, with less than 3 ms overhead per decision.

Ablation Studies

• No single learning modality (RL, imitation, LLM synthesis) is sufficient for sustained high performance; only hybridization exceeds 80% success. The full method (all modalities) achieves the highest rate, indicating the complementary nature of these learning signals.

Discussion and Implications

Theoretical Implications

This capability-centric paradigm imposes an architectural separation-of-concerns, allowing system designers to ensure cognitive continuity, transparency, and auditability—properties essential for deployment in safety-critical or physically situated contexts. The framework generalizes previous modular and skill-learning architectures by introducing explicit lifecycle management, versioning, and runtime-composable skills. These features align closely with software microservices design, but are extended here to handle stochastic, learned policies under formal safety constraints.

In contrast to continual learning approaches that focus on catastrophic forgetting at the parameter level, capability evolution side-steps interference by maintaining strict module isolation. This mitigates the need for complex regularizers (EWC, GEM, etc.) and removes confounding effects on agent identity.

Practical Implications

For long-lived embodied agents (e.g., autonomous robots), system drift, reproducibility failures, and unsafe adaptation are critical risks. By fixing the agent core and evolving only encapsulated modules, continuous improvement is reconcilable with stability guarantees. As skills evolve to adapt to new tasks or stochastic conditions, system behavior remains predictable, auditable, and revertible. Runtime governance provides a domain- and task-independent mechanism for integrating new capabilities without investing safety validation at the learning level.

Limitations

The evaluation is limited to simulated settings, and the generalizability to real-world deployment and additional domains (e.g., navigation, multi-agent settings) requires further investigation. The computational overhead of full lifecycle management versus direct agent adaptation methods is not fully quantified. Overly conservative runtime policies could potentially reject valid behaviors, necessitating more nuanced constraint specification.

Future Directions

Promising extensions include: formalizing agent partial-adaptivity (relaxing strict identity preservation for “soft” cognitive evolution), scaling to distributed multi-agent systems with shared evolving capability registries, and integrating formal type systems for skill input/output verification in complex compositions. The development of standardized benchmarks for the evolutionary improvement of embodied agents, and real-robot evaluations, are important next steps for validating and extending this framework.

Conclusion

This work rigorously formalizes capability evolution as an alternative to agent modification in embodied systems, coupling architectural identity invariance with modular, versioned skill evolution and runtime-governed execution. Empirical data demonstrate that this yields superior success rates, reduced variance, and enforced safety compared to both static, monolithic, and coupled adaptation methods. The paradigm has broad implications for the design of scalable and robust embodied AI, offering a blueprint for future architectures targeting long-term autonomy, safety, and composability in physical environments.