SUN: Persistent Programs for Language-Grounded Control-to-Learning in Long-Horizon Manipulation

Forum topic · 小凯 · 2026-09-02

Summary

This paper introduces Semantically UNified (SUN) Programs, typed executables in which geometric and contact relations are defined once and compiled into aligned Model Predictive Control (MPC) costs, satisfaction predicates, RL rewards, transition guards, and diagnostics. The approach addresses a long-standing gap between model-based control and learned policies in long-horizon robotic manipulation, where existing protocols discard task semantics, requiring hand-crafted rewards and causing behavior drift. The accompanying system, Kuafu, is driven by large vision-language systems to automatically synthesize SUN Programs from language instructions and scene semantics, screens feasibility via MPC, and preserves task semantics during stage-conditioned policy training. Across nine tasks, Kuafu achieves an 82.03% macro success rate, outperforming sparse-reward (35.67%) and Stage-BC (24.75%) baselines. At 8192-way parallelism it generates successful trajectories 10.57x faster than human teleoperation. With 500 trajectories per task, policies trained on Kuafu data (DP3) reach 46.0% simulation success and 34.7% on physical Franka and Kinova robots. The work unifies symbolic planning with data-driven execution without demonstrations or manual dense rewards.

Overview

Field: Machine Learning
Authors: Weiqi Wang, Zhi Li, Yudong Lei
Published: 2025-09-01
arXiv: 2509.00140

Abstract

Bridging model-based control and learned policies in long-horizon manipulation has harbored a silent disagreement: control executes specified objectives, 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.

The Kuafu System

Kuafu is driven by large vision-language systems and:

Automatically synthesizes SUN Programs from language instructions and scene semantics
Screens feasibility via MPC
Retains task semantics while training stage-conditioned policies

Key Results

Success rate: 82.03% macro success rate across nine tasks, versus 35.67% for sparse-reward and 24.75% for Stage-BC baselines
Data throughput: At 8192-way parallelism, generates successful trajectories 10.57x faster than human teleoperation per hour
Policy transfer: With 500 trajectories per task, DP3 policies trained on Kuafu data achieve 46.0% simulation success rate (vs. 22.4% for alternatives) and 34.7% on physical Franka and Kinova robots

Conclusion

These results demonstrate that simulation-screened task semantics can effectively amortize control into robust policies without demonstrations or manual dense rewards, unifying symbolic planning with data-driven execution.

---

*Source: zhichai.net forum post, collected 2026-09-02.*