Research map / Image generation / Fast diffusion sampling
FlowRAM: Grounding Flow Matching Policy with Region-Aware Mamba Framework for Robotic Manipulation
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 FlowRAM, a Mamba-based visuomotor policy that combines a dynamic radius schedule for region-aware 3D point sampling with conditional flow matching for fast keyframe action generation.
Problem
Diffusion-based robotic manipulation policies are slow due to iterative denoising and lack task-relevant region-focused 3D perception, which hurts high-precision manipulation.
Approach
Shrink the 3D perception radius around the noisy end-effector pose over flow-matching time steps (dynamic radius schedule), fuse multimodal tokens with a Mamba block, and generate keyframe poses with conditional flow matching in very few integration steps.
Key results
- Achieves 82.3% average success across 10 multi-task RLBench tasks, the best reported, versus 78.4% for 3D Diffuser Actor.
- Reaches 52.0% average success on 7 high-precision RLBench tasks, versus 40.0% for 3D Diffuser Actor and 39.4% for RVT-2.
- CFM with 2 inference steps keeps competitive performance where DDIM with 2 steps fails; real-world success 49/60 across 6 tasks.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use FlowRAM instead of 3D Diffuser Actor | RLBench multi-task policy · RLBench 10-task multi-task | average success rate: 78.4 -> 82.3 | Helped |
| use FlowRAM instead of 3D Diffuser Actor | high-precision manipulation policy · RLBench 7 high-precision tasks | average success rate: 40.0 -> 52.0 | Helped |
| add dynamic radius scheduling (ablation) | FlowRAM · RLBench multi-task | average success rate: 79.1 -> 82.3 | Helped |
| add the point encoder (ablation) | FlowRAM · RLBench multi-task | average success rate: 81.5 -> 82.3 | Helped |
| replace Mamba with Transformer (ablation) | FlowRAM with DRS and point encoder · RLBench multi-task | average success rate: 82.3 -> 82.1 | No clear effect |
| replace CFM with DDIM (ablation) | FlowRAM with Mamba, DRS, and point encoder · RLBench multi-task | average success rate: 82.3 -> 82.0 | No clear effect |
| replace CFM with DDIM at 2 inference steps (ablation) | FlowRAM · RLBench multi-task | average success rate: 74.2 -> 29.6 | Hurt |
| increase CFM inference steps from 2 to 4 (ablation) | FlowRAM · RLBench multi-task | average success rate: 74.2 -> 77.8 | Helped |
| increase DDIM inference steps from 2 to 4 (ablation) | FlowRAM · RLBench multi-task | average success rate: 29.6 -> 69.8 | Helped |
Surprising findings
- DDIM with only 2 inference steps fails to generate feasible actions, while CFM maintains high success.
- Without dynamic radius scheduling, globally sampled point cloud encoding both increases inference time and reduces performance.
Limitations
- Relies on keyframe-based paradigm and a multi-stage motion planner (BiRRT) to reach predicted poses.
- Real-world evaluation is small (6 tasks, 56 demonstrations).
Experimental setup
Datasets: RLBench High-precision RLBench tasks (7 tasks) Real-world UR5 demonstrations
Benchmarks: RLBench RLBench high-precision tasks real-world UR5 tasks
Metrics: average success rate inference time
Baselines: PerAct GNFactor ManiGaussian Act3D RVT-2 3D Diffuser Actor DDIM DDPM
Builds on: conditional flow matching Mamba PerAct 3D Diffuser Actor Act3D
Closest papers by meaning
- FlowPolicy: Enabling Fast and Robust 3D Flow-Based Policy via Consistency Flow Matching for Robot Manipulation (AAAI 2025 · similarity 0.90)
- DynaFLIP: Rethinking Robotics Perception via Tri-Modal-Dynamics Guided Representation (NeurIPS 2026 · similarity 0.88)
- PDFactor: Learning Tri-Perspective View Policy Diffusion Field for Multi-Task Robotic Manipulation (CVPR 2025 · similarity 0.87)
- Frequency‑Aware Flow Matching for Continuous and Consistent Robotic Action Generation (NeurIPS 2026 · similarity 0.87)
- FreqPolicy: Efficient Flow-based Visuomotor Policy via Frequency Consistency (NeurIPS 2025 · similarity 0.87)
- Generate Subgoal Images before Act: Unlocking the Chain-of-Thought Reasoning in Diffusion Model for Robot Manipulation with Multimodal Prompts (CVPR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.