acceptodds
Withdrawn from ICLR 2027

Closed-Loop Satellite Collision Avoidance: A Benchmark and Planner-Guided Learning Framework

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Satellite collision avoidance (SCA) is a closed-loop sequential decision problem: maneuvers must reduce conjunction risk, limit delivered velocity increment, and recover the spacecraft toward its designated orbit. Existing studies use incompatible assumptions that obscure this trade-off. We introduce SCA-Bench, combining catalog-derived states and historical encounter geometries with finite-burn control, online re-screening, recovery, and single- and multi-object scenarios. Decoupled warning and execution models include measured prediction-execution discrepancy in warning covariance. We also present SCA-Planner, which behavior-clones offline finite-burn planner trajectories before warning-gated proximal policy optimization. On a fixed 128-scenario test set, OrbitZoo obtains the highest collision-probability compliance (99.21%), whereas SCA-Planner uses less cumulative maneuver effort (0.1850 vs. 0.4300 m/s), recovers closer to the designated orbit (0.5800 vs. 0.4600), and attains a higher composite score. Ablations expose distinct operating points; no method dominates safety, maneuver effort, and recovery.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Reinforcement Learning, Imitation Learning, Continuous Control, Benchmark, Satellite Collision Avoidance

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ksXTOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper presents an effort to create a model for collision avoidance SGA-planner. It also includes a benchmark to evaluate collision avoidance SCA-Bench.

Critical strengths

The paper presents an evaluation benchmark for collision avoidance, grounded on real data.

Critical weaknesses

(1) The authors mention that one episode contains 240 intervals ( two hours) for avoidance and recovery. This is not sufficient, and does not resemble what is used in a typical CAM (Collision avoidance manouver), first of all, a manouver is performed in the orbital periods before the TCA (Time of Closest approach), since one period is between 80-128 minutes, for LEO, this does not give enought time to perform a manouvre, and even less to return to the nominal orbital afterwards.

In a non-automated case, CAM are planned with 1-2 days in advance before TCA. Starlink, which has the quickest and more automated process, still plans and performs it in a 12h window before TCA.

(2) It is not clear in which environent this is created.

(3)The test set is limited to 128 cases. The kelvins dataset can only give dataset wide statistics. There are far more collision avoidance cases that can be simulated following the process described in the paper.

(4) The authors should include Basilisk RL in the related work section. The authors fail to identify the novelty of the proposed solution against orbitzoo, which is a general environment that can be used not only for collision avoidance, but also for other tasks, such as orbit transfer or station-docking.

Other small errors:

It is not clear what test128 is. If it is the test set, and there is no other test set, just call it test set.

Foreman missing year L038

Figure 4: some figures are low quality png instead of vectorized images

Incorrect in text citation use L107

Incorrect bibliography:

Review of conjunction probability methods for short-term encounters | request PDF. ResearchGate. URL https://www.researchgate.net/publication/289311065_Review_ of_conjunction_probability_methods_for_short-term_encounters. ( just an example)

Questions for authors

(1) What is the environment used? The code seems incomplete with some imports refering to base_work. What is base_work? (2) Do you consider your new environment to be a novel and relevant part of the paper. Or can your model be used within OrbitZoo or Basilisk RL?

Reviewer_UACQOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces SCA-Bench, a benchmark for closed-loop satellite collision avoidance. It combines catalog-derived spacecraft states with historical encounter geometries and evaluates controllers under multiple aspects. SCA-Planner uses offline planner demonstrations to initialize a continuous action policy and a recurrent intervention gate, followed by phase-aware PPO fine-tuning. Experiments show that SCA-Planner achieves lower maneuver effort and better orbit recovery than the adapted OrbitZoo baseline, but substantially lower collision-probability compliance. The results therefore demonstrate different operating points rather than an established improvement across the safety–effort–recovery trade-off.

Critical strengths

The strongest aspects I think, are the benchmark’s coherent task formulation and the reference learner’s alignment with sparse intervention. Both address practical difficulties in evaluating and learning avoidance behavior.

This benchmark evaluates avoidance and subsequent recovery within a common closed loop, including finite burns and renewed risk screening after maneuvers. Separating prediction from execution and explicitly measuring their discrepancy further strengthens the protocol. In my opinion, this is consequential because a maneuver changes both the spacecraft’s future orbit and its subsequent warning context; isolated maneuver-level evaluation can miss these coupled effects.

Critical weaknesses

My main concerns locate in the benchmark’s reproducibility and the strength of the comparative evidence. These are particularly consequential because the benchmark is the paper’s primary contribution.

  1. Several quantities that directly determine benchmark outcomes are insufficiently specified, including the command-to-acceleration mapping and thrust limits, the collision hard-body radius, and the numerical orbital-tube scales used by ORS. The covariance construction for realized-trajectory collision probability is also unclear. Appendix A.1 reports prediction-error percentiles, but these do not establish calibration of the resulting risk estimates. Consequently, the current description does not adequately support independent reproduction or assessment of the benchmark’s central safety metric.

  2. The comparisons do not establish whether the proposed design improves the attainable trade-off or primarily changes objective preferences. SCA-Planner has 21 noncompliant cases out of 128, versus one for OrbitZoo, so its lower maneuver effort cannot be interpreted independently of its lower compliance. There are no comparisons at matched compliance levels, the demonstration-generating planner is not evaluated as a reference, and each learned configuration uses one training seed. These omissions limit the evidence for the added value of the learning framework beyond producing another operating point.

Questions for authors
  1. What are the exact thrust limits, action scaling, collision radius, and ORS scale vector? How are covariances propagated or assigned when computing realized-trajectory collision probability, and what independent data or checks support calibration of the added model-error covariance?

  2. How does the demonstration-generating planner perform on held-out scenarios, with its computation cost reported? For the optimization baselines, what replanning and recovery mechanisms are used, and what explains Armellin’s compliance matching the no-op baseline despite nonzero maneuver effort?

  3. Under each evaluated policy, how often do secondary warnings and target switches occur, how many are successfully resolved, and how many failures involve secondary objects? Can these outcomes explain the small change in SCA-Planner’s aggregate compliance?

  4. When reward weights or gate thresholds are varied, how do SCA-Planner, the feed-forward-gate variant, and OrbitZoo compare in maneuver effort and recovery at similar compliance levels? Are these conclusions consistent across training seeds?

Reviewer_2PaYOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary
  • Collision avoidance is complicated for satellites due to fuel costs, displacements, and future threats. Therefore collision-avoidance needs to be considered in closed-loop.
  • SCA-Bench is introduced to test different methods. It includes a warning model and an execution model.
  • They pair this with SCA-Planner to demonstrate this benchmark.
  • This setup is nice, though the empirical execution needs a lot of work (e.g. no confidence intervals, no-ops controller doing better than baselines)
  • Unclear if this is the best fit for ICLR
Critical strengths
  • Good, motivated problem and test case
  • Separation of warning model and execution model is nice
  • Separately reporting safety, effort, and recovery is good rather than some aggregate cost
  • The baselines are well selected to balance classical methods and how this platform can also be used for learned approaches.
Critical weaknesses
  • This paper is different from what I typically see in ICLR. It is more of a simulator platform for testing algorithms of any sort. It could certainly be used by the learning community, but I don’t have a good sense on if ICLR is the best venue. The learning algorithm they demonstrate is pretty standard. Perhaps this is a question for the area chair.
  • For ICLR it is more common to provide statistics on the results. Train the policy with multiple seeds, evaluate multiple times, provide confidence intervals
  • No code has been released as of now. It would be good to anonymously provide the code for checking.

Other notes from claude that I thought were useful after checking:

  • There’s no validation set, and some selection appears to use the test set. Appendix B mentions a “recovery-favored checkpoint,” and the reference configuration seems to be chosen partly by CCAS. The paper never describes a validation split, so it’s unclear how checkpoints, the gating threshold, and reward coefficients were chosen without looking at Test128.
  • The learned-method comparison is at single, reward-dependent operating points. For RL methods, where a method lands on the safety-effort-recovery trade-off is largely set by reward weights. OrbitZoo was “adapted” (reward scale, horizon, threshold), but the paper doesn’t say how much tuning it received, how long it trained, or whether it got anything comparable to SCA-Planner’s 1,352 teacher demonstrations. To support “distinct operating points,” both methods should be swept across reward weightings to produce trade-off curves. Otherwise OrbitZoo’s higher CME and lower ORS may just reflect untuned weights
  • The teacher isn’t reported. The offline planner that generates demonstrations replays candidates in the closed-loop environment, which is privileged access. Its Test128 performance is the obvious reference point: is SCA-Planner approaching the teacher, and does PPO actually exceed it anywhere? BC-only at about 64% hazardous resolution suggests a large imitation gap.
  • Table 10 conflicts with Table 3. BC-stage CPCR is 86.67% (tanh) and 93.33% (ReLU) in Table 10, but BC-only is 75.00% in Table 3. CME is 1.3–1.7 m/s in Table 10 vs 0.16 in Table 3. The percentages 86.67 and 93.33 are 13/15 and 14/15, suggesting a separate, much smaller evaluation set. That set isn’t described, and “ReLU improves all four quantities” is a one-episode difference.

The benchmark’s physical realism is limited, and it isn’t positioned against operational practice:

  • The warning threshold of 3×10⁻⁷ is roughly three orders of magnitude below the commonly cited operational 10⁻⁴. A justification is needed.
  • The episode lasts two hours, with TCA at most two hours away. Operational collision avoidance decisions are typically made days in advance.
  • The “truth” model is two-body plus J2, with no atmospheric drag, which is a dominant error source in LEO.
  • The prediction-execution mismatch comes from inverse-SGP4 reconstruction error, not from real orbit-determination or propagation uncertainty.
  • Other objects follow prescribed trajectories.
Questions for authors
  • Why are some of the baselines so bad? Armellin was as safe as doing nothing (29%), and Patera is not much better. Several of the approaches did worse on the composite score compared to doing nothing. It would be good to either refine these or provide an explanation for why these don’t work on your case. For this platform to be useful the baselines should be well-executed and competitive
  • Can you run more seeds and do a statistical analysis and be clear on the validation set?
  • Can you make a strong case for why this paper belongs in ICLR?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.