Research map / Vision-language and multimodal / Robot action and navigation
Trajectory-Level Continuous Action Representation for Robotic Manipulation
Tong Yang, Jingkai Jia, Yuecheng Xu, Xueyao Chen, Chi Zhang, Wenqiang Zhang
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 timestep-level action prediction with CAT latents | SmolVLA · LIBERO | success rate (%): 87.3 -> 90.8 | Improved |
| replace timestep-level action prediction with CAT latents | pi0 · LIBERO | success rate (%): 94.2 -> 96.5 | Improved |
| use CAT instead of FAST or VQ-VLA discretization | SmolVLA · LIBERO | success rate (%): 84.5 -> 90.8 | Improved |
| replace timestep-level prediction with CAT latents | Diffusion Policy (flow matching) · MimicGen 8 tasks | success rate (%): 85.8 -> 86.9 | Improved |
| replace timestep-level prediction with CAT latents | Diffusion Policy (flow matching) · MimicGen Stack 3 | success rate (%): 76 -> 85 | Improved |
| add CAT across control frequencies | SmolVLA · RoboTwin 2.0 at 10Hz | success rate (%): 35.3 -> 45.6 | Improved |
| add CAT across control frequencies | SmolVLA · RoboTwin 2.0 at 16.7Hz | success rate (%): 42.1 -> 46.1 | Improved |
| add CAT | SmolVLA · RoboTwin 2.0 Adjust Bottle (smooth task) | success rate (%): 75.0 -> 76.0 | No clear effect |
| add CAT | SmolVLA · Real-world long-horizon tasks | normalized task score: 45.0 -> 60.5 | Improved |
| use 4 register tokens instead of 2 (ablation) | CAT (SmolVLA) · LIBERO | success rate (%): 88.0 -> 90.8 | Improved |
| use 4 register tokens instead of 8 (ablation) | CAT (SmolVLA) · LIBERO | success rate (%): 87.3 -> 90.8 | Improved |
| add contrastive regularization (lambda 0.1) instead of none (ablation) | CAT (SmolVLA) · LIBERO | success rate (%): 89.3 -> 90.8 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- 4D-VLA — 4D-VLA: Spatiotemporal Vision-Language-Action Pretraining with Cross-Scene Calibration (NeurIPS 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- FreqPolicy: Frequency Autoregressive Visuomotor Policy with Continuous Tokens (NeurIPS 2025 · similarity 0.90)
- Move-Then-Operate: Behavioral Phasing for Human-Like Robotic Manipulation (ICML 2026 · similarity 0.90)
- Chain-of-Action: Trajectory Autoregressive Modeling for Robotic Manipulation (NeurIPS 2025 · similarity 0.90)
- FreqPolicy: Efficient Flow-based Visuomotor Policy via Frequency Consistency (NeurIPS 2025 · similarity 0.89)
- Robots Pre-train Robots: Manipulation-Centric Robotic Representation from Large-Scale Robot Datasets (ICLR 2025 · similarity 0.89)
- Beyond World-Frame Action Heads: Motion-Centric Action Frames for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.88)
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.
