OrbitTAMP: Grounding Language Models for Task and Motion Planning in Spacecraft Rendezvous
Spacecraft rendezvous and proximity operations currently require specialized engineers to translate high-level mission intent into dynamically feasible trajectories, creating a scalability bottleneck for distributed space systems. This paper introduces OrbitTAMP, a hierarchical task-and-motion planning framework that leverages pretrained LLMs to map natural-language commands into a partial mission specification, then systematically completes unspecified behavior, timing, and waypoint decisions via a domain-specific behavior graph and planning modules. Final trajectory optimization ensures dynamical feasibility and safety, aiming to provide an auditable, scalable language interface for spacecraft operations.
The paper presents a well-motivated hierarchy that preserves operator intent while resolving underspecification through structured, domain-grounded planning modules rather than LLM hallucination. Empirically, the scaffolded architecture achieves 98% exact recovery of partial mission specifications with a frontier LLM and far outperforms direct LLM generation baselines that frequently exhaust token limits or violate physical constraints. However, the demonstration is limited to a narrow LEO inspection case study with only two-to-three-phase missions and a small hand-engineered behavior graph, leaving scalability to richer operational vocabularies and closed-loop execution largely unproven.
The stage-wise decomposition—intent parsing, behavior-duration planning, waypoint generation, and trajectory optimization—is conceptually clean and empirically effective. Crucially, the label-first benchmark design constructs commands from known mission specifications without any LLM involvement in the reference creation, eliminating circular evaluation and providing an unambiguous ground truth. The controlled ablations across architectural scaffolds, model backends, and test-time compute allocations are carefully executed and yield actionable takeaways, such as the complementary roles of verifier-guided revision and downstream plan search.
A major limitation is the constrained operational envelope: the case study supports only five waypoint domains and thirteen behavior primitives, with missions restricted to $M\in\{2,3\}$ phases in a single LEO scenario. This narrow scope raises questions about whether the behavior graph can scale without requiring prohibitive expert engineering. Additionally, the offline leg library used for fast candidate screening assumes circular Keplerian dynamics with no $J_2$ perturbations, an acknowledged source of deployment error that weakens the fidelity of feasibility predictions before full trajectory optimization.
The evidence convincingly supports the superiority of hierarchical scaffolding over direct LLM generation. The unscaffolded baseline failed frequently—exhausting its token budget on a substantial fraction of commands—whereas the structured architecture reliably produced intent-aligned trajectories. Comparisons among Qwen3.5-9B, Ministral3-8B, and GPT-5.6 Terra show that the structured interface itself provides strong grounding, though frontier models still dominate exact-match recovery. The deterministic rule-based parser provides a useful sanity check, exceeding compact LLMs on familiar linguistic patterns but collapsing on an unseen terse register, thereby validating the need for learned semantic parsing.
The paper documents dynamics equations, the SCP formulation, behavior tables, and the MissionConfig schema in detail, and delegates hyperparameters and network architectures to supplementary material. However, the authors do not state whether code, prompts, or the procedural command-generation templates will be released; without these artifacts, independent reproduction would require substantial reimplementation. The reliance on supplementary sections for key model and algorithmic details means reproducibility depends heavily on the availability and completeness of those additional documents.
Spacecraft rendezvous and proximity operations (RPO) are currently planned through an expertise-intensive process in which engineers translate high-level operational intent into safe, dynamically feasible trajectories, creating a bottleneck to scalable operations. Large language model (LLM)-based agents could offer an intuitive interface for this process, although their outputs are not inherently grounded in orbital dynamics, operational constraints, or the structure of admissible spacecraft maneuvers. To exploit their semantic reasoning while ensuring the generated plan's physical validity, this paper presents a hierarchical framework for spacecraft task-and-motion planning (TAMP) that grounds LLM reasoning in a graph of reusable behaviors and domain-specific planning modules. Within this framework, a pretrained LLM maps a natural-language command to a partial mission specification. The associated planners then resolve unspecified decisions within the admissible operational space. Finally, trajectory optimization converts the completed mission specification into a dynamically feasible trajectory. Numerical experiments demonstrate that this architecture substantially improves intent recovery over direct LLM generation, achieving 98% exact recovery of partial mission specifications across all evaluated splits when backed by frontier LLMs. Additional test-time-compute experiments show that, for a compact 9B model, verifier-guided revision increases exact recovery from 75% to 88%, while broader behavior-plan search independently improves selection among admissible trajectory realizations. Overall, these results establish a scalable and auditable foundation for language-driven agentic planning of spacecraft RPO.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.