Evaluating Robustness and Adaptability in Learning-Based Mission Planning for Active Debris Removal

Autonomous mission planning for Active Debris Removal (ADR) must balance efficiency, adaptability, and strict feasibility constraints on fuel and mission duration. This work compares three planners for the constrained multi-debris rendezvous problem in Low Earth Orbit: a nominal Masked Proximal Policy Optimization (PPO) policy trained under fixed mission parameters, a domain-randomized Masked PPO policy trained across varying mission constraints for improved robustness, and a plain Monte Carlo Tree Search (MCTS) baseline. Evaluations are conducted in a high-fidelity orbital simulation with refueling, realistic transfer dynamics, and randomized debris fields across 300 test cases in nominal, reduced fuel, and reduced mission time scenarios. Results show that nominal PPO achieves top performance when conditions match training but degrades sharply under distributional shift, while domain-randomized PPO exhibits improved adaptability with only moderate loss in nominal performance. MCTS consistently handles constraint changes best due to online replanning but incurs orders-of-magnitude higher computation time. The findings underline a trade-off between the speed of learned policies and the adaptability of search-based methods, and suggest that combining training-time diversity with online planning could be a promising path for future resilient ADR mission planners.

💡 Research Summary

This paper investigates the robustness and adaptability of three autonomous mission planners for Active Debris Removal (ADR) in Low Earth Orbit. The task is a constrained sequential decision‑making problem: a service spacecraft must visit and rendezvous with multiple debris objects while respecting a limited Δv budget, a total mission duration, and feasibility constraints that depend on remaining resources and debris visitation status. The authors compare (1) a nominal Masked Proximal Policy Optimization (PPO) agent trained on a fixed mission configuration (7‑day duration, 3 km/s Δv), (2) a domain‑randomized Masked PPO agent that is exposed during training to a range of mission durations (3–7 days) and Δv budgets (1–3 km/s), and (3) a plain Monte Carlo Tree Search (MCTS) planner that replans online at every decision step using a UCT selection rule and 200 rollouts per step. All three methods share the same discrete action space (transfer to any unvisited debris or to a refueling station) and enforce feasibility through action masking, ensuring that illegal maneuvers never influence learning or search.

The experimental platform is a custom Gymnasium‑compatible environment (SpaceDebrisStressTestEnv) that simulates realistic co‑elliptic Hohmann transfers, safety‑ellipse final approach, and optional mid‑mission refueling. Fifty debris objects are randomly generated for each episode, and each planner is evaluated on 100 independent test cases for three scenarios: (a) nominal (7 days, 3 km/s), (b) reduced fuel (Δv = 1 km/s), and (c) reduced mission time (3 days). Performance is measured by the number of successful debris rendezvous; computation time is also recorded.

Results show a clear trade‑off. In the nominal scenario, the nominal PPO achieves the highest average debris count (29.1 ± 1.1), slightly outperforming the domain‑randomized PPO (28.2 ± 1.2) and MCTS (27.1 ± 0.9). When the mission duration is shortened, the domain‑randomized PPO adapts best, achieving 14.1 ± 0.7 visits versus 12.6 ± 0.6 for nominal PPO and 11.9 ± 0.3 for MCTS. Under severe fuel limitation, the nominal PPO collapses to 3.2 ± 0.9 visits, while the domain‑randomized PPO improves to 8.1 ± 2.9, yet both are outperformed by MCTS, which reaches 15.0 ± 0.4 visits. Computationally, both PPO variants require less than one second per episode on a consumer‑grade CPU, whereas MCTS needs roughly four minutes per episode due to repeated environment cloning and rollouts.

The analysis highlights that learned policies provide ultra‑fast inference suitable for onboard deployment but are vulnerable to distributional shift when mission constraints differ from training conditions. Domain randomization mitigates this vulnerability by exposing the policy to a spectrum of constraints, thereby improving adaptability with only modest loss in nominal performance. MCTS, by contrast, is inherently robust to constraint changes because it replans online, but its computational burden makes it impractical for real‑time spacecraft hardware.

The authors conclude that a hybrid approach—combining the speed of a robust, domain‑randomized policy with limited, targeted online search (e.g., shallow MCTS or Monte‑Carlo rollouts guided by the policy)—could offer a promising path forward. Future work is suggested on meta‑learning for rapid adaptation, adaptive action‑masking mechanisms, and hardware‑in‑the‑loop validation on actual satellite platforms. This study provides a quantitative baseline for the trade‑offs between learning‑based and search‑based planners in constrained ADR missions, informing designers when to prioritize inference speed versus adaptability.