SUN: Persistent Programs For Language-Grounded Control-to-Learning-to-Real Policies

The provided search context does not contain information regarding a system or framework explicitly named "SUN" or "SUN Programs" for language-grounded control.

However, the context describes several related approaches to long-horizon manipulation and geometric reasoning:

• Traversal-as-Policy: This approach externalizes long-horizon agent control as executable tree traversal in a log-distilled Gated Behavior Tree (GBT), creating a persistent, verifiable policy artifact independent of model weights. It uses pre-execution gates to ensure deterministic safety.
• Hierarchical RL–MPC Framework: This method separates geometric reasoning (handled by a high-level RL policy predicting contact intentions) from contact dynamics (handled by a low-level MPC). This decoupling enables efficient sim-to-real transfer and data efficiency for contact-rich tasks.
• Code as Policies: This paradigm uses Language Model Programs for embodied control, where LLMs synthesize symbolic plans or executable action scripts that low-level controllers instantiate, allowing for flexible task specification via natural language.
• Knowledge-Guided Manipulation: This framework uses Multi-Task Reinforcement Learning with a dynamic-relation mechanism and graph neural encoders to incorporate spatial, containment, and affordance knowledge into the control loop.

SUN introduces a persistent-program approach to language-grounded manipulation in which task semantics are specified once as typed executables and then compiled into multiple downstream control artifacts. Rather than treating natural-language task descriptions, rewards, or demonstrations as the primary executable substrate, SUN Programs encode geometric and contact relations—such as object poses, approach constraints, contact modes, and task dependencies—as first-class, checkable program objects. These programs can then be compiled into aligned model predictive controllers and learned policies, so that the same task specification informs both analytic control and data-driven behavior across the control-to-learning-to-real pipeline.

The key contribution is a shared executable representation that preserves verified task semantics as policies move from specification to simulation to deployment. By making contact and geometric constraints explicit and typed, SUN can reduce the semantic drift that often occurs when learned policies are trained on proxies for the intended task, such as handcrafted rewards, imperfect demonstrations, or loosely grounded language prompts. This alignment is especially important for long-horizon manipulation, where small inconsistencies in contact assumptions, object relations, or task ordering can compound over many steps and make failures difficult to diagnose.

The work matters because it addresses a central gap in language-conditioned robotics: natural language can express intent, but reliable physical execution requires precise, verifiable structure and control interfaces. SUN positions persistent programs as the bridge between language grounding, model-based control, and learned policies, enabling task semantics to remain interpretable, checkable, and reusable across different embodiments and environments. For a curated collection, the material is notable for combining programmatic task specification, MPC, and learned control under a single language-grounded framework for manipulation.