Force sensing is a crucial modality for Vision-Language-Action (VLA) frameworks, as it enables fine-grained perception and dexterous manipulation in contact-rich tasks. We present Force-Distilled VLA (FD-VLA), a novel framework that integrates force awareness into contact-rich manipulation without relying on physical force sensors. The core of our approach is a Force Distillation Module (FDM), which distills force by mapping a learnable query token, conditioned on visual observations and robot states, into a predicted force token aligned with the latent representation of actual force signals. During inference, this distilled force token is injected into the pretrained VLM, enabling force-aware reasoning while preserving the integrity of its vision-language semantics. This design provides two key benefits: first, it allows practical deployment across a wide range of robots that lack expensive or fragile force-torque sensors, thereby reducing hardware cost and complexity; second, the FDM introduces an additional force-vision-state fusion prior to the VLM, which improves cross-modal alignment and enhances perception-action robustness in contact-rich scenarios. Surprisingly, our physical experiments show that the distilled force token outperforms direct sensor force measurements as well as other baselines, which highlights the effectiveness of this force-distilled VLA approach.

An autonomous bin-picking system for grasping various cluttered packages can significantly benefit logistics by reducing manual labor and streamlining processing. We propose a bin-picking system that includes a novel multi-mode hybrid gripper combining suction and pinch, and a corresponding vision-based grasp planning strategy based on unseen object instance segmentation. The system was evaluated in simulation achieving a 71.4% success rate, compared to suction (53.9%) and Hand-E (39.3%). Real-world experiments further validated its practicality in logistics scenarios.

Generative visuomotor policies rely heavily on the conditioning representation to guide the synthesis of accurate and stable control sequences. Yet, standard visual encoders produce high-dimensional embeddings that often lose fine-grained spatial information while retaining substantial redundancy. To address this bottleneck, we propose Gaussian Spotlight, designed to construct a latent spatial keypoint embedding that serves as a more precise and manipulation-aware conditioning signal. Gaussian Spotlight first generates a state-conditioned anisotropic Gaussian Attention Field that selectively amplifies spatial regions critical for interaction. It then transforms these enhanced regions into implicit, latent keypoint embeddings via an attention-guided skip-layer aggregation pathway. Extensive experiments across diverse real-world manipulation tasks and simulation benchmarks demonstrate that Gaussian Spotlight consistently enhances policy performance.

Real-time efficiency is critical for visuomotor policy learning, as any delay in action generation can accumulate over sequential control steps. In this work, we introduce 3D-LOT Policy, a latent prototype-guided optimal transport flow-matching framework for effective single-step action generation. Our approach encodes 3D observations into a compact latent space that preserves task-relevant spatial information and induces prototype structures to serve as anchors for policy learning. Our experiments demonstrate that 3D-LOT achieves lower latency while maintaining or even surpassing baseline performance, offering a practical solution for fast and robust visuomotor policy learning.