Research map / Reinforcement learning / Visual and representation RL
Falcon: Fast Visuomotor Policies via Partial Denoising
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.
Falcon is a training-free plug-in that accelerates diffusion visuomotor policies by reusing partially denoised actions from historical steps, selected via one-step Tweedie estimates and a thresholded temperature softmax, achieving 2-7x fewer sampling steps.
Problem
Diffusion policies need many sampling steps for action generation, which hurts real-time inference; existing acceleration methods either need retraining or degrade performance at low step counts.
Approach
Exploit sequential dependencies between consecutive actions: store partially denoised action sequences from past decision steps in a latent buffer, use one-step Tweedie estimates conditioned on the current observation to find those consistent with the current reference action, and start denoising from a sampled partial action instead of Gaussian noise.
Key results
- On RoboMimic, DDIM+Falcon reduces NFE from 16 to about 6.5-10.5 steps while keeping success rates close to DDIM; DPMSolver+Falcon achieves about 2-7x speedup (e.g., 7x on Lift ph).
- Against step-matched reduced DPMSolver*, Falcon retains performance; DPMSolver* drops 10.6% on Square ph and 35.6% on Square mh.
- In real-world dexterous grasping, Falcon+DDIM... specifically Falcon+DDPM achieved 3.07x speedup over DDPM (30.98 vs 100 NFE) with 100% success, and 2.86x on square stick insertion with 90% success matching DDPM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Falcon to DDPM | DDPM · RoboMimic Lift ph | NFE: 100.0±0.0 -> 12.9±0.3 | Helped |
| Add Falcon to DDIM | DDIM · RoboMimic Lift ph | NFE: 16.0±0.0 -> 7.5±0.2 | Helped |
| Add Falcon to DPMSolver | DPMSolver · RoboMimic Square mh | success rate: 0.84±0.36 -> 0.90±0.30 | Helped |
| Add Falcon to DDPM | DDPM · Dexterous Grasping | speedup: +3.07x speedup over DDPM | Helped |
| Add Falcon to DDPM | DDPM · Square Stick Insertion | success rate: 90% -> 90% | No clear effect |
| Use a very small epsilon threshold (ablation) | Falcon on Square_ph · RoboMimic Square ph | score | No clear effect |
| Use a large epsilon threshold (ablation) | Falcon on Square_ph · RoboMimic Square ph | score | Hurt |
| Increase exploration rate δ from zero (ablation) | Falcon on Transport_ph · RoboMimic Transport ph | score | Helped |
| Increase exploration rate δ further (ablation) | Falcon on Transport_ph · RoboMimic Transport ph | speedup | Hurt |
| Replace adaptive selection with a fixed noise level K/2 or K/5 (ablation) | Falcon · Falcon selection mechanism evaluation | score | Hurt |
| Use reduced-step DPMSolver without Falcon (ablation) | DPMSolver · RoboMimic Square ph | success rate: 10.6% drop compared to DPMSolver(16 steps)+Falcon | Hurt |
| Use reduced-step DPMSolver without Falcon (ablation) | DPMSolver · RoboMimic Square mh | success rate: 35.6% drop compared to DPMSolver(16 steps)+Falcon | Hurt |
Surprising findings
- Falcon's speedup is smaller on complex tasks (Square, Transport, ToolHang) because the reference action differs more between steps, so low-noise starting points are rarely selected.
- Falcon-accelerated DDPM on PushT preserves multimodal reaching trajectories while Consistency Policy collapses to one mode.
Limitations
- Hyperparameters (threshold epsilon, exploration rate delta, k_min, buffer size) must be tuned per task, as the conclusion notes.
- Speedup is smaller on complex, high-variation tasks where historical partial actions are less consistent with the current action.
Experimental setup
Datasets: RoboMimic Franka Kitchen Push-T Multimodal Block Pushing MetaWorld ManiSkill2
Benchmarks: RoboMimic MetaWorld Franka Kitchen Push-T Multimodal Block Pushing ManiSkill2 real-world dexterous grasping real-world square stick insertion
Metrics: success rate NFE speedup runtime memory usage
Baselines: DDPM DDIM DPMSolver SDP Consistency Policy 3D Diffusion Policy
Builds on: DDPM DDIM DPMSolver Diffusion Policy Streaming Diffusion Policy
Closest papers by meaning
- One-Step Diffusion Policy: Fast Visuomotor Policies via Diffusion Distillation (ICML 2025 · similarity 0.88)
- FreqPolicy: Efficient Flow-based Visuomotor Policy via Frequency Consistency (NeurIPS 2025 · similarity 0.88)
- Latent Policy Barrier: Learning Robust Visuomotor Policies by Staying In-Distribution (NeurIPS 2025 · similarity 0.87)
- Two-Steps Diffusion Policy for Robotic Manipulation via Genetic Denoising (NeurIPS 2025 · similarity 0.87)
- FreqPolicy: Frequency Autoregressive Visuomotor Policy with Continuous Tokens (NeurIPS 2025 · similarity 0.86)
- DynaMo: In-Domain Dynamics Pretraining for Visuo-Motor Control (NeurIPS 2024 · similarity 0.86)
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.