Research map / Reinforcement learning / Visual and representation RL

ForceFlow: Learning to Feel and Act via Contact-Driven Flow Matching

NeurIPS 2026 · Reinforcement learning › Visual and representation RL · approach: Dexterous manipulation transfer · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes ForceFlow, a flow-matching DiT policy that fuses temporal force history asymmetrically (AdaLN for force/proprio, cross-attention for vision), jointly predicts next-step force, and hands over to a VLM pointing stage (V2F) for OOD spatial generalization, evaluated on six real-world contact-rich tasks.

Problem

Vision-centric imitation learning policies fail at contact-rich manipulation because high-dimensional visual features overshadow low-dimensional force signals (modal masking), and they do not separately generalize physical interaction and spatial layout.

Approach

Asymmetric multimodal fusion that conditions a flow-matching DiT on force-history and proprioception via AdaLN while using vision only via cross-attention, jointly predicting motion and next-step contact force, plus a VLM pointing-based V2F handover that separates approach from force-driven interaction.

Key results

• ForceFlow achieves 81.67% average success rate over six real-world tasks vs. 45% for ForceVLA (best baseline).
• ForceFlow reduces average force MAE to 8.23 N vs. 23.31 N for ForceVLA.
• Without force history (1-step), Stamp success drops from 85% to 55%; removing both history and prediction gives 40%.

Surprising findings

• Removing vision entirely during the contact stage (w/o vis) collapses success to 0% on Plug, Insert, and Clean Vase, showing vision remains necessary during interaction.
• ForceVLA + V2F still achieves only 13.33% average success on spatial OOD tasks, indicating end-to-end baselines fail even after spatial alignment.

Limitations

• Relies on high-fidelity force/torque sensors, limiting deployment on low-cost platforms.
• V2F switching is fixed and triggered by positional criteria rather than adaptive.
• Small evaluation with 20 trials per task and a single robot platform.

Experimental setup

Datasets: expert teleoperation demonstrations (6 real-world tasks) manually annotated VQA pointing pairs

Benchmarks: six real-world contact-rich tasks (Stamp, Plug, Press, Insert, Clean Whiteboard, Clean Vase)

Metrics: success rate MAE force cost (N)

Baselines: π0.5 ACT Diffusion Policy ForceVLA ForceFlow (w/o Force)

Builds on: Flow Matching ForceVLA Diffusion Policy Embodied-R1

Closest papers by meaning

• ForceVLA: Enhancing VLA Models with a Force-aware MoE for Contact-rich Manipulation (NeurIPS 2025 · similarity 0.90)
• Touch in the Wild: Learning Fine-Grained Manipulation with a Portable Visuo-Tactile Gripper (NeurIPS 2025 · similarity 0.88)
• FlowPolicy: Enabling Fast and Robust 3D Flow-Based Policy via Consistency Flow Matching for Robot Manipulation (AAAI 2025 · similarity 0.87)
• FlowHOI: Flow-based Semantics-Grounded Generation of Hand-Object Interactions for Dexterous Robot Manipulation (NeurIPS 2026 · similarity 0.87)
• DynaFLIP: Rethinking Robotics Perception via Tri-Modal-Dynamics Guided Representation (NeurIPS 2026 · similarity 0.87)
• FlowRAM: Grounding Flow Matching Policy with Region-Aware Mamba Framework for Robotic Manipulation (CVPR 2025 · similarity 0.86)