SUN and Kuafu: Persistent Programs for Language-Grounded Robot Control
This post from the zhichai.net forum introduces SUN (Semantically UNified Programs), a framework described in the paper *"SUN: Persistent Programs For Language-Grounded Control-to-Learning-to-Real Policies"* (arXiv:2608.31167), together with its implementation system Kuafu (named after the sun-chasing giant of Chinese mythology).
Key points
The problem: a silent disagreement between control and learning
- Control systems demand precise, hand-written instructions for every step—they are rigid but verifiable.
- Learning systems need massive demonstration data—they are flexible but data-hungry.
- Both approaches discard task semantics: controllers only track joint angles and trajectories; learned policies only imitate action sequences without understanding why. This forces hand-crafted reward functions and lets behavior drift away from control-verified paths.
- MPC cost functions
- Satisfaction predicates
- RL rewards
- Transition guards
- Diagnostics for human debugging
- 82.03% macro success rate across 9 manipulation tasks, vs. 35.67% (sparse reward baseline) and 24.75% (Stage-BC).
- At 8192-way parallel scale, Kuafu generates the equivalent of 10.57 hours of human teleoperation of successful trajectories per hour—reducing dependence on costly human demonstrations.
- Data transfers to simulation training: DP3 policies reach 46.0% success (vs. 22.4% for alternatives).
- Sim-to-real: 34.7% success on physical Franka and Kinova robots—notable given the well-known sim-to-real gap.
- Wang, W., Li, Z., Lei, Y., et al. (2026). *SUN: Persistent Programs For Language-Grounded Control-to-Learning-to-Real Policies*. arXiv:2608.31167.
SUN's core idea: define once, compile everywhere
SUN programs are typed executables encoding geometric relations (above, beside, contact) and contact relations (grasp, place, push). Defined once, they automatically compile into:
The author uses a recipe analogy: one "tomato-scrambled-eggs" recipe can serve a beginner (detailed MPC), an expert (RL shorthand), a robot (joint trajectories), and QA (success checklists).
Kuafu's four-stage pipeline
1. Language understanding — a vision-language model parses "put the red block inside the blue box" into action, objects, and relations. 2. Program synthesis — a structured SUN program with stages (Approach, Grasp, Transport, Release), each with guards and a satisfaction predicate. 3. Feasibility screening — MPC "rehearses" the plan, checking collisions, joint limits, and actuator forces; failures yield diagnostics like "target unreachable." 4. Policy learning — a stage-conditioned policy is trained, retaining semantics rather than blindly imitating motions.
Reported results
Why semantics matter
The post argues modern end-to-end AI suffers a kind of "amnesia": it can act but not explain. Like a trained dog versus a taught child, an RL agent may find statistically optimal but semantically nonsensical strategies (e.g., pushing objects with the arm's side instead of "grasping"). Explicit semantics make systems:
1. Interpretable — you can ask which stage failed and on which object. 2. Debuggable — inspect a single failing stage without retraining everything. 3. Composable — combine grasp/transport/place primitives like Lego bricks into complex tasks.
Outlook
The vision is robots that understand intent behind language rather than matching keywords—a home robot told "I want coffee" that reasons about locating beans, cleaning the grinder, filling the kettle, and retrieving a cup. The author closes with Feynman's principle: if you can't explain a task clearly enough for a six-year-old (which block? which box? how do you know it's done?), you haven't truly understood it—and SUN forces exactly that clarity.