Research map / Reinforcement learning / Visual and representation RL
Real-Time Robot Execution with Masked Action Chunking
Haoxuan Wang, Gengyu Zhang, Yan Yan, Yuzhang Shang, Ramana Rao Kompella, Gaowen Liu
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 |
|---|---|---|---|
| add LoRA alone without masked training (ablation) | pi0/flow policy · Kinetix | solve rate | No clear effect |
| progressively add masked chunking and prefix-preserved sampling components (ablation) | REMAC · Kinetix | solve rate | Improved |
| use self-conditioned curriculum schedule instead of pure ground-truth or pure self-conditioned inputs (ablation) | REMAC · Kinetix | solve rate and stability | Improved |
| add BID at test time (ablation) | REMAC · Kinetix d=4 | solve rate: 0.779 -> 0.781 | Improved |
| add RTC at test time (ablation) | REMAC · Kinetix d=4 | solve rate: 0.779 -> 0.791 | Improved |
| increase q_min / q_max (ablation) | REMAC · Kinetix | solve rate | No clear effect |
| add learned mask embedding (ablation) | REMAC · Kinetix | solve rate | No clear effect |
| reduce training data from 200 to 30 trajectories (ablation) | REMAC · real robot Grasp-Hard | completion progress: 0.820 -> 0.810 | No clear effect |
| use REMAC vs RTC | REMAC · Kinetix d=4 | solve rate: 0.588 -> 0.760 | Improved |
| add REMAC to ACT vs naive async | ACT · Transfer Cube d=4 | success rate: 0.40 -> 0.74 | Improved |
| add REMAC to ACT | ACT · Insert Box d=10 | success rate: 0.14 -> 0.16 | Mixed |
| corrupt delay input with noisy and spiky values | REMAC · Kinetix d=3 | solve rate: 0.817 -> 0.718 (d=3 spiky) | Hurt |
Closest papers by meaning
- Real-Time Execution of Action Chunking Flow Policies (NeurIPS 2025 · similarity 0.93)
- PACE: Phase-Aware Chunk Execution for Robot Policies with Action Chunking (NeurIPS 2026 · similarity 0.92)
- FASTER: Rethinking Real-Time Flow VLAs (NeurIPS 2026 · similarity 0.89)
- Bidirectional Decoding: Improving Action Chunking via Guided Test-Time Sampling (ICLR 2025 · similarity 0.89)
- Improving Generative Behavior Cloning via Self-Guidance and Adaptive Chunking (NeurIPS 2025 · similarity 0.86)
- Chain-of-Action: Trajectory Autoregressive Modeling for Robotic Manipulation (NeurIPS 2025 · 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.
