Research map / Reinforcement learning / Visual and representation RL
CARP: Visuomotor Policy Learning via Coarse-to-Fine Autoregressive Prediction
Zhefei Gong, Pengxiang Ding, Shangke Lyu, Siteng Huang, Mingyang Sun, Wei Zhao, Zhaoxin Fan, Donglin Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace Diffusion Policy with CARP next-scale autoregressive prediction | Diffusion Policy · Robomimic state-based Square | inference time for 400 actions (s): 35.21 -> 3.07 | Improved |
| replace next-token autoregressive BET with CARP | BET · Robomimic state-based Square | success rate: 0.54 -> 0.98 | Improved |
| use CARP instead of diffusion policy on image-based Square | CARP · Robomimic image-based Square | success rate: 0.92 -> 0.88 | Hurt |
| use CARP instead of SDP in multi-task visual setting | SDP · MimicGen 8 tasks | average success rate: 0.78 -> 0.85 | Improved |
| use CARP instead of SDP on Hammer Cleanup | SDP · MimicGen Hammer Cleanup | success rate: 1.00 -> 0.98 | Hurt |
| add EMA during training (ablation) | CARP · training stability and performance | success rate: 4-5% improvement | Improved |
| increase number of scales K above 4 (ablation) | CARP · Can, Square, Kitchen | success rate | No clear effect |
| use fewer scales (K below 4) (ablation) | CARP · Can, Square, Kitchen | success rate | Hurt |
| use cosine similarity instead of Euclidean distance in quantizer (ablation) | CARP tokenizer · training stability | training stability | Improved |
| use separate VQ-VAE per action dimension instead of one joint VQ-VAE (ablation) | CARP tokenizer · training stability and token interpretability | stability | Improved |
| use CARP instead of Diffusion Policy on real robots (RTX 2060) | Diffusion Policy · real-world Cup and Bowl | inference speed and success rate: about 8x faster, up to 10% higher success | Improved |
| use CARP instead of VQ-BeT on Kitchen p4 | VQ-BeT · Franka Kitchen | p4 success rate: 0.71 -> 0.98 | Improved |
Closest papers by meaning
- FreqPolicy: Frequency Autoregressive Visuomotor Policy with Continuous Tokens (NeurIPS 2025 · similarity 0.92)
- Chain-of-Action: Trajectory Autoregressive Modeling for Robotic Manipulation (NeurIPS 2025 · similarity 0.91)
- Latent Policy Barrier: Learning Robust Visuomotor Policies by Staying In-Distribution (NeurIPS 2025 · similarity 0.90)
- Falcon: Fast Visuomotor Policies via Partial Denoising (ICML 2025 · similarity 0.90)
- FreqPolicy: Efficient Flow-based Visuomotor Policy via Frequency Consistency (NeurIPS 2025 · similarity 0.90)
- One-Step Diffusion Policy: Fast Visuomotor Policies via Diffusion Distillation (ICML 2025 · similarity 0.90)
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.
