Research map / Reinforcement learning / Visual and representation RL
FreqPolicy: Frequency Autoregressive Visuomotor Policy with Continuous Tokens
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 FreqPolicy, a frequency-domain autoregressive visuomotor policy that applies DCT to action sequences, generates coarse-to-fine low-to-high frequency levels, and uses continuous tokens with a diffusion loss (building on MAR/FAR).
Problem
Visuomotor policies must produce precise actions while remaining computationally efficient; existing diffusion methods are slow, and autoregressive methods discretize continuous action spaces and ignore the hierarchical structure of motion.
Approach
Represent action sequences via DCT and perform autoregressive generation from low- to high-frequency levels, with each step producing a continuous token that conditions a diffusion head, plus frequency-aware masking inherited from FAR.
Key results
- Average success rate of 67.9% across 48 simulation tasks in Adroit, DexArt and MetaWorld, vs. 64.6% for DP3 and 65.8% for Mamba Policy.
- Inference of 0.21s vs. 2.11s for Diffusion Policy on Robomimic/Push-T settings, with comparable success rates.
- Real-world handover runs at 70 FPS with one iteration vs. 25 FPS for DP3.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use hierarchical frequency-domain DCT action modeling instead of omitting DCT (ablation) | FreqPolicy · Adroit, DexArt, Meta-World (10 tasks) | average success rate (%): 61.4 -> 71.5 | Helped |
| add frequency-aware masking (ablation) | FreqPolicy · Adroit tasks: Hammer, Door, Pen, Pick Out of Hole, Soccer, Stick Pull | average success rate: 40 -> 58 | Helped |
| increase prediction horizon to 64 (action step 32) (ablation) | FreqPolicy · Adroit tasks: Hammer, Door, Pen, Pick Out of Hole, Soccer, Stick Pull | average success rate: 58 -> 44 | Hurt |
| use prediction horizons 8 or 16 rather than smaller or larger alternatives (ablation) | FreqPolicy · Adroit tasks: Hammer, Door, Pen, Pick Out of Hole, Soccer, Stick Pull | average success rate: 53; 58; 58; 56; 50 | No clear effect |
| use FreqPolicy instead of Mamba Policy* | visuomotor policy · Adroit, DexArt, Meta-World (48 tasks) | average success rate (%): 65.8 ± 14.4 -> 67.9 ± 14.2 | Helped |
| use FreqPolicy instead of DP3* | visuomotor policy · Adroit, DexArt, Meta-World (10 tasks) | average success rate (%): 67.8 -> 71.5 | Helped |
| use FreqPolicy instead of DP3 | visuomotor policy · RoboTwin, Block Hammer Beat and Block Handover | success rate (%): 47.7 ± 4.0 -> 42.0 ± 4.2; 86.0 ± 1.0 -> 80.7 ± 9.7 | Hurt |
| use FreqPolicy instead of DP-C | visuomotor policy · Robomimic and Push-T | inference time (s): 2.11 -> 0.21 | Helped |
| use FreqPolicy instead of DP3 | visuomotor policy · Adroit handover (real world) | inference speed (FPS): 25 FPS -> 70 FPS | Helped |
| train FreqPolicy with Gaussian noise at std=0.1 instead of high-quality demonstrations | FreqPolicy · Adroit tasks, average | success rate (%): 33% versus 19-23% for baselines | Mixed |
| use FreqPolicy instead of Diffusion Policy | visuomotor policy · RoboCasa (24 multitask atomic tasks) | success rate (%): 25.6% -> 27.4% | Helped |
| use FreqPolicy instead of GR00T-N1 | visuomotor policy · RoboCasa (24 multitask atomic tasks) | success rate (%): 32.1% -> 27.4% | Hurt |
Surprising findings
- Removing the frequency-aware mask drops average success on the ablation tasks from 58 to 40, showing masking is critical.
- Low-frequency-only reconstruction (about 12%) retains stable performance on Robomimic tasks, while Adroit tasks degrade sharply without high frequencies.
Limitations
- Performance on 2D tasks has room for improvement, and it decreases when frequency partitioning is too fine-grained.
- Only condition inputs consistent with DP/DP3 were explored; alternative conditioning was not studied.
- VLA multitask results on RoboCasa (27.4%) trail GR00T-N1 (32.1%).
Experimental setup
Datasets: Adroit DexArt MetaWorld Robomimic Push-T RoboTwin RoboCasa
Benchmarks: Adroit DexArt MetaWorld Robomimic Push-T RoboTwin RoboCasa
Metrics: success rate inference time
Baselines: Diffusion Policy DP3 Mamba Policy BeT CARP GR00T-N1 BC-Transformer
Builds on: MAR FAR Diffusion Policy 3D Diffusion Policy FAST
Closest papers by meaning
- FreqPolicy: Efficient Flow-based Visuomotor Policy via Frequency Consistency (NeurIPS 2025 · similarity 0.94)
- Latent Policy Barrier: Learning Robust Visuomotor Policies by Staying In-Distribution (NeurIPS 2025 · similarity 0.90)
- Chain-of-Action: Trajectory Autoregressive Modeling for Robotic Manipulation (NeurIPS 2025 · similarity 0.89)
- DynaMo: In-Domain Dynamics Pretraining for Visuo-Motor Control (NeurIPS 2024 · similarity 0.88)
- Fourier Controller Networks for Real-Time Decision-Making in Embodied Learning (ICML 2024 · similarity 0.87)
- Robots Pre-train Robots: Manipulation-Centric Robotic Representation from Large-Scale Robot Datasets (ICLR 2025 · 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.