Factor-Aware Mixture-of-Experts with Pretrained Encoder for Combinatorial Generalization
Authors: Feihong Zhang, Guojian Zhan, Zeyu He, Yinuo Wang, Likun Wang, Tianze Zhu, Yao Lyu, Tao Zhang, +3 more
Organizations: Tsinghua University, Beijing, China · SunRisingAI Ltd., Beijing, China · EFORT Intelligent Robot Co., Ltd., Anhui, China
Abstract
The integration of pretrained encoders with diffusion policies has become a dominant paradigm for visual robotic manipulation. However, it still struggles to generalize across complex environments with varying factors such as lighting and surface textures. To address this, we propose FAME, a framework that integrates a factor-aware mixture-of-experts (MoE) with a pretrained encoder to enhance generalization to environmental variations. FAME follows a three-stage training process: (1) policy warmup, where a diffusion policy is trained on standard-environment data with a frozen encoder; (2) factor-specific adapter training, where lightweight adapters inserted between the frozen encoder and the temporarily frozen policy are trained on customized datasets, each targeting a distinct environmental variation; and (3) joint fine-tuning, where a central router and the warmed policy are trained on mixed data to handle multiple factors jointly. FAME is ``factor-aware'' because the central router softly weights frozen factor-specific adapters as a dense MoE, enabling combinatorial generalization across multiple factors. Evaluations on the Meta-World benchmark show that FAME outperforms diffusion policy baselines by 34%. We further validate FAME in a real-world pick-and-place task using a compact model trained on newly collected data, where FAME achieves a 35% improvement in generalization under real-world variations.
Diffusion-based visuomotor policies operating directly in raw action spaces conflate scene comprehension with trajectory generation within a single denoising process. The resulting velocity field must simultaneously encode scene information and generate precise trajectories, increasing learning complexity and limiting performance on tasks demanding precise temporal coordination across multiple arms. To simplify this joint learning problem, we introduce Latent Diffusion Policy (LDP), a two-stage framework performing flow matching in a deliberately shaped latent space. By absorbing scene understanding into an observation-conditioned CVAE encoder, LDP concentrates the conditional distribution of each observation. Consequently, the flow model avoids implicitly resolving scene-dependent structures; instead, it generates within a pre-concentrated distribution featuring a smoother velocity field, simplifying learning from limited demonstrations. Furthermore, to capture temporal dependencies among latent tokens, LDP trains with per-token diffusion forcing and employs staircase inference sampling to resolve the resulting distributional mismatch. We also propose reconstruction FID (rFID) as a lightweight proxy predicting downstream task success solely from latent space statistics. On coordination-intensive tasks from RoboTwin 2.0, LDP outperforms DP3 by a substantial margin and transfers effectively to real-world bimanual deployments.
Diffusion-based policies have established a new standard for precise robotic manipulation but face a critical scalability bottleneck: high-performance models are computationally expensive, while lightweight alternatives often fail to generalize across diverse multi-task environments. Mixture-of-Experts (MoE) architectures offer a promising path to efficiency by activating only a subset of parameters. However, existing MoE routing mechanisms typically rely on low-level noise or latent statistics, ignoring the compositional nature of manipulation tasks. This can fragment reusable behaviors across experts, limiting interpretability and transferability. We introduce Semantically Structured Mixture-of-Experts Diffusion Policy (SMoDP) for compositional robotic manipulation, a framework that grounds expert specialization in semantic task structure. SMoDP leverages a lightweight, inference-time skill predictor, supervised by offline annotations from Vision-Language Models (VLMs), to route action chunks to experts specialized for specific behavioral phases. To ensure robust assignment, we propose a dual contrastive alignment strategy that grounds multi-modal observations in language-defined skill semantics (Inter-modal) while enforcing routing consistency across visually distinct but functionally related behaviors (Intra-modal). Our approach outperforms representative diffusion and MoE-based baselines on multi-task benchmarks with significantly improved parameter efficiency and demonstrates effective compositional transfer to novel tasks through parameter-efficient fine-tuning. Project website: https://deng-cy20.github.io/SMoDP/
Robotic tasks are typically specified by a tuple of factors, such as the object to be grasped, the obstacles to be avoided, the color of the target, and so on. Collecting expert demonstrations for every combination of factor values grows combinatorially. We present factored diffusion policies: a single shared diffusion network trained with per-factor null-token dropout, whose score decomposes additively across factors at inference. Under approximate conditional independence between factors given the action-observation pair, this composition approximates the true joint score with a bounded uniform error, reducing the training-task budget from a product of factor cardinalities to a sum. A trajectory-tube certificate chains this score-level bound through the reverse-time sampling ODE and a contracting tracking controller into a closed-loop state-trajectory tube whose radius factors into an ODE-sensitivity constant and a per-factor score-error budget. Unlike compositional-diffusion methods for control that combine separately trained networks, we use one shared network. Drone racing experiments confirm both the generalization bound and the certificate. On state-based multi-gate racing, the factored policy passes 90% of held-out gates -- matching an oracle -- while a K-network composition baseline collapses to 3%; on vision-based single-gate traversal, it transfers zero-shot to an unseen venue with +11.7pp success-rate gain and 2.4X crash-rate reduction.