Paper Overview
- Field: Machine Learning (ML)
- Authors: Weiqi Wang, Zhi Li, Yudong Lei
- Published: 2025-09-01
- arXiv: 2509.00140
- Automatically synthesizes SUN Programs from natural language and scene semantics
- Screens feasibility via MPC
- Retains semantics while training stage-conditioned policies
- At 8192-way parallelism, Kuafu generates successful trajectories 10.57x faster per hour than human teleoperation.
- With 500 trajectories per task, DP3 policies trained on Kuafu data reach 46.0% simulated success (vs. 22.4% for alternatives) and 34.7% on physical Franka and Kinova robots.
Summary
Bridging model-based control and learned policies in long-horizon manipulation has harbored a silent disagreement: control executes specified objectives, while learning amortizes that behavior into a reactive policy. Yet existing protocols discard task semantics, leaving rewards hand-crafted and behavior drifting from what control verified.
The authors introduce Semantically UNified (SUN) Programs — typed executables where geometric and contact relations are defined once and compiled into aligned Model Predictive Control (MPC) costs, satisfaction predicates, RL rewards, transition guards, and diagnostics.
Kuafu, the proposed system, is driven by large vision-language systems and:
Results
Across nine tasks:
| Setting | Success Rate | |---|---| | Kuafu | 82.03% | | Sparse reward baseline | 35.67% | | Stage-BC baseline | 24.75% |
Conclusion
These results demonstrate that simulation-screened task semantics can effectively amortize control into robust policies without demonstrations or manually designed dense rewards, unifying symbolic planning with data-driven execution.
---
*Collected automatically on 2026-09-02.*