PaperlayerView exact revision on arXiv

Source-linked AI summary

OrbitTAMP: Grounding Language Models for Task and Motion Planning in Spacecraft Rendezvous

Yuji Takubo, Daniele Gammelli, Marco Pavone, Simone D'Amico

arXiv:2610.01093v1cs.ROcs.AImath.OC

TL;DR

Spacecraft RPO planning requires translating high-level intent into safe, dynamically feasible trajectories, while existing interfaces leave a gap in preserving partial intent and resolving unspecified decisions. OrbitTAMP grounds language-based planning in spacecraft behavior abstractions and domain-specific modules, and its experiments show improvements in semantic reliability and trajectory quality through test-time computation.

  • Problem

    Existing spacecraft planning interfaces depend on expert-specified maneuvers or complete trajectory representations, while language-grounded TAMP generally does not preserve partial RPO intent as behavior, timing, and waypoint decisions are completed.

    Source: A. Spacecraft Planning, Language… · p. 3, B. Learning-Guided Task and Motion… · p. 3

  • Method

    The framework represents commands as partial mission specifications and progressively completes them through admissible spacecraft behaviors, planning modules, and trajectory verification.

    Source: §VI · p. 15

  • Results

    Test-time computation improves semantic reliability and realized trajectory quality, with verifier-guided revision affecting intent parsing and broader planning search improving trajectory selection.

    Source: D. RQ3: Test-Time Compute Across the… · pp. 14–15, §VI · p. 15

  • Takeaways & Limitations

    The framework is a step toward scalable, auditable human–spacecraft interfaces for language-driven RPO planning.

    Source: §VI · p. 15

Abstract

from arXiv · show

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.

II. Related Work

Prior spacecraft-planning and language-grounded TAMP methods offer complementary capabilities, but leave a gap in preserving partial operator intent while completing and verifying spacecraft plans. This work unifies these directions around partial mission specifications and admissible spacecraft behaviors.

A. Spacecraft Planning, Language Grounding, and Trajectory Optimization

  • Prior spacecraft planning and trajectory-optimization approaches leave no structured mechanism to preserve partial intent, resolve unspecified decisions, and verify behavior under orbital dynamics and safety constraints.Their interfaces depend on expert-specified maneuvers or complete trajectory-level representations.Source: A. Spacecraft Planning, Language… · p. 3

B. Learning-Guided Task and Motion Planning

  • Language-grounded TAMP methods generally require a sufficiently specified symbolic problem or skill sequence, unlike RPO planning with unresolved behavior, timing, and waypoint decisions.Source: B. Learning-Guided Task and Motion… · p. 3
  • The framework represents language as a partial mission specification and completes its unspecified decisions through admissible spacecraft behaviors, trajectory-aware evaluation, and trajectory optimization.Source: B. Learning-Guided Task and Motion… · p. 3

III. Methodology

  • The planning objective is to infer admissible behavior sequences, solve for continuous trajectories, and resolve missing information without overwriting stated intent.The shared MissionConfig evolves through partial and completed specifications across the pipeline.Source: §III · p. 3

A. Common Planning Representation

The framework uses a behavior–waypoint graph and a structured MissionConfig to represent admissible spacecraft maneuvers and preserve both specified and unresolved mission decisions. Planning modules complete behavior, duration, and waypoint fields while retaining operator-provided information and checking the completed specification through trajectory optimization.

A. Common Planning Representation

  • The behavior–waypoint graph represents admissible behavior transitions between waypoint domains, making reusable spacecraft behaviors the basis for structuring the planning search.Each graph edge pairs a behavior with waypoint domains and an allowable transfer-time range.Source: A. Common Planning Representation · p. 4

2. Mission Configuration Interface

  • MissionConfig stores operator priorities and constraints alongside planning fields for behaviors, durations, total flight time, and waypoints, with unresolved entries marked ⊥.Its alternative field types preserve explicit intent while allowing downstream completion of unspecified decisions.Source: §2 · p. 4

B. Hierarchical Planning

  • Planning modules progressively complete the shared mission specification while preserving populated fields; trajectory optimization verifies the result against dynamical feasibility and safety requirements.The completion interfaces allow deterministic or learning-based implementations.Source: B. Hierarchical Planning · p. 4

2. Behavior-Duration Sequence Planning

  • The behavior-duration planner enumerates graph-valid plans consistent with the initial specification, filters candidates using estimated convergence, and ranks the remainder by operator preferences.The evaluator estimates trajectory-optimization convergence and mission metrics, avoiding a separate full-horizon solve for every candidate.Source: §2 · p. 5

3. Waypoint Generation

Waypoint generation completes the selected behavior-duration plan with continuous waypoint states while reconciling operator-specified waypoint information. The resulting mission specification is passed to trajectory optimization, where waypoints can be fixed or allowed to vary within their domains.

3. Waypoint Generation

  • Waypoint generation proposes continuous relative-state waypoints within the domains associated with the selected behavior sequence, using an implementation-specific placement policy.The policy may use direct optimization, a geometric or physics-informed heuristic, or a learned predictor.Source: §3 · p. 5
  • The waypoint module preserves operator-specified coordinates, accepting proposals within symbolic domains and completing only unspecified coordinates in partial relative states.Proposed waypoints may later serve as fixed constraints or warm-start references.Source: §3 · p. 5
  • Trajectory optimization enforces dynamics, safety, and waypoint conditions; its variable contract allows waypoints to move within selected domains, whereas its fixed contract requires passage through proposed waypoints.Source: §3 · pp. 5–6

IV. Case Study: On-Orbit Inspection in Low Earth Orbit

The case study evaluates an on-orbit inspection scenario in low Earth orbit and ranks plans using four preference metrics. Its setup varies the keep-out-zone radius and target orbital phase while considering two- or three-phase behavior sequences.

IV. Case Study: On-Orbit Inspection in Low Earth Orbit

  • The inspection case study uses two- or three-phase behavior sequences and varies the keep-out-zone radius and target’s initial mean anomaly.The radius takes values 30, 40, or 50 m, and the initial mean anomaly takes five values from 0° to 288°.Source: §IV · p. 6

A. Preference Metrics

  • Plans are ranked on fuel, flight time, effective observation exposure, and safety margin, with each metric assigned a preferred optimization direction.Fuel and flight time are minimized; effective observation exposure and safety margin are maximized.Source: A. Preference Metrics · p. 6

B. Behavior-Waypoint Graph

The behavior–waypoint graph encodes admissible transfers between waypoint domains and their duration ranges. The case-study trajectory optimizer then solves a minimum-fuel problem with waypoint and safety constraints.

B. Behavior-Waypoint Graph

  • The graph has five waypoint domains and 13 behavior-primitive types, yielding 21 directed edges and 466 graph-valid two- or three-phase paths.The waypoint space constrains drift and separates E/I-separated domains from flat domains; admissible transfers and durations are enumerated in Tables I and II.Source: B. Behavior-Waypoint Graph · p. 6, Table I · p. 7, Table II · p. 7
  • The graph-valid paths and discrete duration choices produce 49,350 behavior-duration plans before scenario variation and continuous waypoint domains are included.This count combines 89 paths with two phases and 377 paths with three phases.Source: B. Behavior-Waypoint Graph · p. 7

C. Trajectory Optimization

  • Trajectory optimization uses ROE dynamics with the secular J2 effect and imposes passive safety constraints at controlled-trajectory nodes and post-failure drift evaluation times.The safety constraint is evaluated over a one-orbit horizon sampled at 50 equally spaced timesteps.Source: C. Trajectory Optimization · p. 7
  • Chance constraints and continuous-time safety constraints are possible extensions, but are outside the scope of this case study.The stated safety formulation evaluates prescribed trajectory nodes and sampled post-failure drift times.Source: C. Trajectory Optimization · p. 7

D. Behavior-Duration Planning and Waypoint Generation

The planning modules realize behavior-duration and waypoint choices through deterministic or learned alternatives, while an LLM conservatively parses commands into partial mission specifications. Downstream planning completes unspecified decisions while retaining operator-provided information.

2. Heuristic Waypoint Generation

  • The heuristic waypoint policy selects domain-center waypoints, choosing the closest connected E/I-separated component while preserving the previous waypoint’s sign combination.For flat domains, it selects the domain center.Source: §1 · p. 7, §2 · p. 8

1. Coupled Behavior–Waypoint Realization

  • The offline leg library assumes a circular Keplerian orbit without J2 perturbations, an approximation that could induce deployment error.The assumption reduces the library to 2,160 entries across the three keep-out-zone radii at the nominal semimajor axis.Source: §1 · p. 7
  • The coupled realization screens behavior-duration candidates using online first-leg solves and offline leg-library entries, then composes leg metrics to rank feasible waypoint chains.Fuel, flight time, and exposure are summed across legs; safety margin is the minimum leg-level value, and a separate full-horizon optimization generates the final verified trajectory.Source: §1 · pp. 7–8

D. Behavior-Duration Planning and Waypoint Generation

  • The learned waypoint policy is supervised on converged optimization trajectories and parameterized to keep decoded waypoints within graph-prescribed domains, including disconnected domains.A learned behavior-duration surrogate predicts convergence and preference metrics, filters candidates below a convergence threshold, and ranks the rest with the deterministic preference map.Source: §2 · p. 8

E. Intent Parsing

  • The intent parser fills only fields explicitly supported by an operator command and leaves unspecified fields blank, preserving them for downstream planning.Commands are generated label-first for three 100-command splits; the held-out structural split uses an unseen terse radio register within the supported mission families and ontology.Source: E. Intent Parsing · p. 8

V. Results

The Results section evaluates architectural scaffolding, model capability, and test-time computation, and first validates waypoint and behavior-duration planning modules.

V. Results

  • The evaluation tests architectural scaffolding, model capability, and test-time computation, asking how they affect intent recognition, failure modes, mission satisfaction, and trajectory preference.The three research questions are assessed at both individual-module and complete language-to-trajectory pipeline levels.Source: §V · pp. 8–9

A. Module-Level Validation and Representative End-to-End Realization

  • Under variable-waypoint SCP, both structured proposals achieved approximately 95% success and substantially lower fuel than random; under fixed waypoints, the minimum-displacement heuristic had the highest success rate.For the fixed contract, the learned proposal yielded lower fuel among converged cases but a lower success rate; the minimum-displacement heuristic with variable-waypoint SCP was adopted subsequently.Source: A. Module-Level Validation and… · p. 9, Table III · p. 10

2. Behavior-Duration Planning

  • Behavior-duration selectors are compared with an offline best-of-64 random-search reference, while Qψ uses the variable-waypoint configuration corresponding to its training policy.The 100 scenarios balance four lexicographic priority orders, with fuel, flight time, observation, and safety margin each prioritized in 25 scenarios.Source: §2 · pp. 9–10
  • Mean candidate percentiles were 98.9% and 99.6% for leg-library selection under strict and baseline margins, versus 96.2% and 96.5% for Qψ; both maintained near-universal SCP convergence.Figure 3 compares preference win rates against best-of-K graph-valid candidates; the best-of-64 reference requires 64 full-horizon SCP evaluations per scenario.Source: Figure 3 · p. 10, §2 · p. 10

3. Representative End-to-End Realization

The experiments trace how commands become completed mission specifications and trajectories, then compare structured scaffolding and language-model capability across the evaluated splits.

3. Representative End-to-End Realization

  • The representative command preserves its ordered maneuvers, one-orbit near-target phase, total-time bound, and fuel priority in M0 while leaving phase durations and continuous waypoints unspecified.The planner completes the sequence with durations [2.5, 1.0, 1.5] orbits within the five-orbit bound; full-horizon SCP then adjusts waypoints while enforcing dynamics and safety constraints.Source: §3 · p. 10, Table V · p. 11

B. RQ1: Effects of Structured Scaffolding

  • Direct M2 receives the MissionConfig schema and behavior–waypoint graph but bypasses behavior-duration search, while the unscaffolded baseline receives neither representation.Both direct-generation baselines may use up to 3,000 completion tokens, compared with 900 per call for the hierarchical architecture.Source: B. RQ1: Effects of Structured… · pp. 10–11
  • The unscaffolded planner exhausts its 3,000-token limit on 45, 44, and 69 commands across the development, held-out-instance, and held-out-structural splits, respectively.The structured architecture recovers intent alignment, and more often realizes the preferred trajectory than Direct M2 when both are intent-aligned.Source: B. RQ1: Effects of Structured… · p. 12

C. RQ2: Effect of LLM Capability

  • GPT-5.6 Terra achieves 98% exact-M0 recovery on all three splits, while Qwen3.5-9B has lower exact recovery and greater sensitivity to the structural split.Compact models nevertheless show high atomic precision and recall, indicating that most operator-specified fields are recovered even when the full specification does not match exactly.Source: C. RQ2: Effect of LLM Capability · p. 12
  • The rule-based parser achieves 87% reference-M⋆ satisfaction and 92% pipeline completion on development and instance splits, but 60% and 85%, respectively, on the structural split.The development and instance splits share semantic frames, whereas the structural split changes linguistic register and phrasing.Source: C. RQ2: Effect of LLM Capability · pp. 12–13

D. RQ3: Test-Time Compute Across the Planning Hierarchy

Test-time computation is allocated either to semantic parsing or downstream behavior planning. Revision improves exact intent recovery, while broader planning search improves selection among trajectories for a fixed parsed specification.

D. RQ3: Test-Time Compute Across the Planning Hierarchy

  • Verifier-guided revision checks schema, ontology, structural constraints, and selected inferable movement patterns; detected correctable inconsistencies trigger feedback to the same LLM.Nrev is a maximum revision budget, so additional computation is invoked only when the verifier detects a correctable inconsistency.Source: D. RQ3: Test-Time Compute Across the… · p. 13
  • Exact-M0 recovery increases from 75% without revision to 88% at Nrev = 2, with no further improvement through Nrev = 5.The remaining errors are primarily semantically incorrect but structurally valid interpretations that the verifier cannot observe.Source: D. RQ3: Test-Time Compute Across the… · pp. 13–14
  • With Nhyp = 3, exact recovery reaches 76% without revision and 87% with Nrev = 2; Nhyp = 5 adds no gain in the selected specification despite increased inference cost.The candidate pool's coverage of the exact reference specification continues to increase with more hypotheses, shifting the constraint toward aggregation and selection.Source: D. RQ3: Test-Time Compute Across the… · pp. 14–15
  • For a fixed parsed specification, the planner ranks graph-valid behavior-duration candidates, optimizes the top Nplan candidates, and selects among successful trajectories by realized preference.This search uses additional trajectory-optimization compute rather than additional LLM calls.Source: D. RQ3: Test-Time Compute Across the… · p. 15
  • Evaluating the top five behavior-duration candidates yields a preferred solution for 34% of commands at Nrev = 0 and 43% at Nrev = 2, relative to evaluating only the top-ranked candidate.The benefit is complementary to semantic refinement: planning search improves trajectory quality after semantic grounding.Source: D. RQ3: Test-Time Compute Across the… · p. 15

VI. Conclusion

The conclusion presents a hierarchical spacecraft TAMP architecture that grounds LLM reasoning in spacecraft behaviors, operational constraints, and orbital dynamics. It reports experimental benefits and identifies multi-spacecraft coordination and online replanning as future work.

VI. Conclusion

  • The architecture grounds LLM reasoning in spacecraft behavior abstractions, operational constraints, and orbital dynamics to generate RPO trajectories.It uses a structured mission representation, reusable behavior graph, and domain-specific planning modules to progressively complete a verifiable mission specification.Source: §VI · p. 15
  • The hierarchy separates semantic grounding, completion of unspecified mission decisions, and physical trajectory verification while preserving operator-provided information.The resulting trajectory satisfies the adopted dynamics and safety constraints.Source: §VI · p. 15
  • Numerical experiments report advantages over one-shot LLM planning and characterize how model capability and test-time computation relate to intent recognition and trajectory quality.Future work will extend the framework to multi-spacecraft coordination and closed-loop execution with online replanning.Source: §VI · p. 15

A. Learning-based Waypoint Generation

Waypoint generation is trained by imitation from optimized, graph-valid mission samples, while behavior-duration planning combines language parsing with learned candidate evaluation. The two modules use separate training objectives and constraints to complete different parts of the mission specification.

A. Learning-based Waypoint Generation

  • 20,000 waypoint-training rows are generated from randomized graph-valid behavior-duration sequences, scenarios, and waypoint samples, with converged optimized sequences serving as imitation targets.Waypoints are sampled on an 11-point grid, and the rows are split 80%/10%/10% for training, validation, and testing.Source: A. Learning-based Waypoint Generation · p. 18
  • The waypoint MLP has three 256-unit hidden layers and represents disconnected coordinates with magnitude logits and binary signs.Its inverse affine-sigmoid map keeps predictions inside graph-prescribed waypoint domains.Source: A. Learning-based Waypoint Generation · p. 18
  • Waypoint imitation training uses Gaussian negative log-likelihood for continuous targets and binary cross-entropy for sign targets, with equal weight for two- and three-phase examples.Training is supervised on converged rows.Source: A. Learning-based Waypoint Generation · p. 18

B. Learning-based Behavior-Duration Planning

  • The behavior-duration metric surrogate combines an MLP and GRU, predicting optimization convergence and mission metrics; convergence uses all rows while metric loss uses converged rows.Safety-margin values are clipped at 25 m before scaling, and convergence probabilities are calibrated on the validation set.Source: B. Learning-based Behavior-Duration… · p. 19
  • At deployment, graph-valid behavior-duration candidates consistent with M0 are evaluated in a batch, filtered below predicted convergence probability 0.5, and ranked by ρP.The selected candidate is the top-ranked remaining behavior-duration plan.Source: B. Learning-based Behavior-Duration… · p. 19
  • Parsing rules preserve only stated intent, phase order, and supplied ontology, leaving omitted values unspecified rather than completing the mission.If verification fails, revision feedback includes deterministic errors while retaining the prior response context; revisions must preserve unaffected fields and avoid adding unstated intent.Source: B. Learning-based Behavior-Duration… · p. 19

D. Command-Benchmark Construction

The command benchmark renders the same mission semantics in multiple linguistic styles, then varies mission-instance sampling and language style across its evaluation splits.

D. Command-Benchmark Construction

  • Four command styles—formal, natural, concise, and radio—render the same reference mission intent using different linguistic realizations.The styles preserve the supported mission semantics and behavior–waypoint ontology.Source: D. Command-Benchmark Construction · p. 19
  • The development and held-out-instance splits use formal, natural, and concise commands, with independently sampled mission specifications in the held-out-instance split.The development split uses development samples.Source: D. Command-Benchmark Construction · p. 19
  • The held-out-structural split uses independently sampled mission specifications rendered exclusively in the unseen radio register to introduce a linguistic-structure shift.Table II summarizes the styles used across the benchmark splits.Source: D. Command-Benchmark Construction · p. 19

E. Verification of M0

The M0 verification process checks required structure and consistency, then applies targeted checks to stated movement sequences and behavior–waypoint reachability.

E. Verification of M0

  • M0 must contain required MissionConfig fields and data types, use the paper’s defined ontology, and have valid, consistent time bounds.Relative-state waypoints require six coordinates, and behavior, duration, and waypoint sequences must have matching phase counts.Source: E. Verification of M0 · pp. 19–20
  • A deterministic movement check covers ordered waypoint-region lists, approach followed by departure, and direct cross-side flybys or fast passages.At zero revision budget, disagreement rejects M0; with revisions, feedback is returned to the LLM, while other command patterns receive no movement-specific feedback.Source: E. Verification of M0 · p. 20
  • A final graph check resets an inconsistent behavior to unspecified when another allowed behavior reaches the requested waypoint, preserving that waypoint without completing the sequence.The check begins from the known initial region and does not fill unspecified fields.Source: E. Verification of M0 · p. 20
Loading 2610.01093v1…