Technical Summary: Emergent Compositional Skills in Mixture-of-Experts VLAs

Problem Statement

Vision-language-action models (VLAs) have demonstrated significant promise in generalizing across diverse robotic manipulation tasks. However, they are typically trained and deployed as monolithic policies. This architecture presents limitations: it is difficult to identify reusable skills, compose behaviors hierarchically, or adapt specific parts of a policy without retraining the entire model. In contrast, many robotic tasks naturally decompose into reusable behavioral modes (e.g., reaching, grasping, placing). The central question of this work is whether a modular policy structure, explicitly learning such skills, can be derived directly from data without pre-specified task decomposition, hierarchy, or manually defined skill libraries.

Methodology

The authors propose training a Mixture-of-Experts (MoE) action policy atop existing pretrained VLA backbones, specifically $\pi_0$ and SmolVLA. Both backbones utilize flow matching for action generation. The approach is defined by two primary design choices:

1. LoRA Mixture-of-Experts: The MoE architecture replaces only the Feed-Forward Network (FFN) sublayer of each action layer, while self-attention remains shared across all experts. Each expert is realized as a low-rank (LoRA) delta ($r=16$) applied to the base FFN. The policy is formulated as:
   $$\text{FFN}_\ell(x) = \text{FFN}^{\text{base}}_\ell(x) + \sum_{e \in R} w_e \left( \text{FFN}^{(e)}_\ell(x) - \text{FFN}^{\text{base}}_\ell(x) \right)$$
   Here, $R$ is the set of routed experts with renormalized weights $w_e$. The LoRA matrices $B$ are zero-initialized, ensuring the policy exactly reproduces the pretrained backbone at step zero, allowing specialization to emerge around a strong prior rather than from scratch.

2. Whole Forward Pass Routing: Unlike standard MoE transformers that route per token per layer, this approach performs routing once per policy forward pass. A small MLP router $g_\phi$ takes a context vector encoding visual, linguistic, and proprioceptive states and emits logits over $E$ experts. A single selection (top-$k$) is applied uniformly across all action-expert layers. This ensures that each expert functions as a coherent end-to-end behavior or "skill" rather than a collection of independent layer-wise decisions.

Training Objective: The model is trained using the backbone's flow-matching behavior-cloning loss combined with an auxiliary load-balancing term ($\lambda_{LB} = 0.01$) to prevent routing collapse onto a single expert.

Key Results and Analysis

The authors evaluate the approach on the LIBERO-10 benchmark, addressing four core questions regarding the nature of the learned skills:

• Emergence of Qualitatively Distinct Skills: The router autonomously learns to select experts that correspond to distinct low-level behaviors. Qualitative analysis reveals experts specializing in specific phases of manipulation, such as "transporting a grasped object," "releasing and retracting," or "approaching and grasping thin handles." These skills are reused across structurally different tasks (e.g., placing a mug vs. a moka pot), suggesting the experts encode phase-level skills rather than task-specific policies.
• Reusable vs. Task-Specific Experts: Analysis of selection frequencies reveals a dual role for experts. Some experts (e.g., Expert A and B) are reused frequently (35–45% selection rate) across multiple tasks, acting as general primitives. Others are highly specialized, activating almost exclusively for specific tasks (e.g., Expert C on Task T5), effectively handling idiosyncratic edge cases and keeping the reusable experts "crisp."
• Composition of Primitives: The router acts as an implicit high-level sequencer, stitching these primitives into longer-horizon trajectories. For example, a task might involve Expert C (grasp) followed by Expert A (place). Notably, even in failure cases, the policy exhibits interpretable behavior, repeatedly invoking known primitives (e.g., attempting to grasp and release multiple times) rather than degenerating into arbitrary noise.
• Manual Routing Validation: To verify that the skills are intrinsic to the experts and not artifacts of the router's policy, the authors performed manual routing experiments. Substituting the router's choice with a specific expert (e.g., forcing a "grasp" expert in a context where it was not originally selected) resulted in the expected qualitative behavior. This confirms that the experts possess reusable, generalizable capabilities independent of the router's default selection.
• Performance: The MoE policy achieves task performance comparable to a dense, monolithic baseline finetuned from the same checkpoint. This demonstrates that skill specialization can be achieved without sacrificing task competence.

Significance and Claims

The paper claims that hierarchical, skill-based learning for robots does not necessarily require additional machinery such as high-level planners, options frameworks, or pre-trained skill libraries. Instead, a simple MoE architecture applied to standard imitation learning can yield:

1. Emergent Compositionality: The model learns to decompose tasks into a small set of modular, reusable primitives.
2. Interpretability: Failures and successes can be understood as sequences of known primitives, making the policy more transparent.
3. Data-Driven Modularity: The compositional structure emerges solely from demonstration data, driven by the interaction between a router and specialized experts.

The authors acknowledge limitations, noting that some learned skills remain task-specific or uninterpretable, and the mapping between experts and primitives is not always perfectly precise. However, the work represents a step toward modular, interpretable robot policies that emerge from data alone.