HiMaCon: Discovering Hierarchical Manipulation Concepts from Unlabeled Multi-Modal Data

TL;DR

This work tackles generalization in robotic manipulation by learning hierarchical manipulation concepts from unlabeled multi-modal demonstrations. It introduces a self-supervised framework that jointly optimizes a Cross-Modal Correlation Network and a Multi-Horizon Future Predictor to learn continuous concept latents $\mathbf{z}$ that capture cross-modal invariants and multi-scale temporal structure, formalized through an objective that emphasizes conditional mutual information and horizon-aware prediction. These manipulation concepts are integrated into imitation learning via joint prediction of actions and concepts, improving transfer to novel objects, barriers, and real-world variations, while producing interpretable clusters that align with human manipulation primitives. Empirical results on LIBERO benchmarks and real-robot deployments demonstrate robust gains over baselines, with analyses revealing both cross-modal grounding and hierarchical organization as key drivers of improved generalization and recovery in complex tasks. The approach offers a practical and interpretable pathway to more adaptable robotic systems that can reason about high-level sub-goals while executing precise low-level actions.

Abstract

Effective generalization in robotic manipulation requires representations that capture invariant patterns of interaction across environments and tasks. We present a self-supervised framework for learning hierarchical manipulation concepts that encode these invariant patterns through cross-modal sensory correlations and multi-level temporal abstractions without requiring human annotation. Our approach combines a cross-modal correlation network that identifies persistent patterns across sensory modalities with a multi-horizon predictor that organizes representations hierarchically across temporal scales. Manipulation concepts learned through this dual structure enable policies to focus on transferable relational patterns while maintaining awareness of both immediate actions and longer-term goals. Empirical evaluation across simulated benchmarks and real-world deployments demonstrates significant performance improvements with our concept-enhanced policies. Analysis reveals that the learned concepts resemble human-interpretable manipulation primitives despite receiving no semantic supervision. This work advances both the understanding of representation learning for manipulation and provides a practical approach to enhancing robotic performance in complex scenarios.

Figures (13)

• Figure 1: Manipulation concepts enhance generalization. Top: Training data with cups and containers without barriers. Middle: Without manipulation concepts, policies fail when encountering barriers. Bottom: With our concept enhancement, policies adapt accordingly.
• Figure 2: The proposed self-supervised manipulation concept discovery and policy enhancement.Stage 1: The concept encoder ($\mathcal{E}$) processes multi-modal robot demonstrations to extract concept latents. These latents are refined through two objectives: (1) the Cross-Modal Correlation Network ($\mathcal{C}$) employs a mask-and-predict strategy to capture persistent patterns across sensing modalities (Sec. \ref{['subsec:multi-modal']}); (2) the Multi-Horizon Future Predictor ($\mathcal{F}$) enables concept latents to organize hierarchically into multi-horizon sub-goals based on coherence thresholds ($\epsilon$) (Sec. \ref{['subsec:multi-hierarchy']}). Stage 2: The learned concepts are integrated into policy learning through a backbone network ($\pi_h$) with concept ($\pi_z$) and action ($\pi_a$) prediction heads, regularizing action generation with structured manipulation knowledge (Eq. \ref{['eq:policy_align_mc']}).
• Figure 3: Conditional mutual information between modality pairs. Values conditioned on concept latents from our method versus the All baseline that does not model cross-modal correlations. A: agentview, H: eye-in-hand vision, P: proprioception.
• Figure 3: Multi-granular task decomposition through concept latent clustering. Visualization of sub-processes derived by clustering manipulation concept latents at different coherence thresholds ($\epsilon$) for the task "open the top drawer and put the bowl in it." Higher $\epsilon$ values (top rows) produce coarser decompositions, while lower values (bottom rows) yield finer-grained segmentation. The emergent sub-processes naturally align with semantic task components, for example, the third segment in row 2 corresponds to "put bowl in drawer," while the second segment in row 4 corresponds to "pull drawer open." This demonstrates our method's ability to discover hierarchical, human-interpretable task structures without explicit supervision.
• Figure 4: Semantic alignment of learned concepts. Cosine similarity between concept latents grouped by human-defined sub-goals. Diagonal patterns demonstrate that our approach discovers concepts that exhibit clustering patterns corresponding to meaningful manipulation primitives.