Action Chunking with Transformers for Image-Based Spacecraft Guidance and Control

Alejandro Posadas-Nava; Andrea Scorsoglio; Luca Ghilardi; Richard Linares; Roberto Furfaro

REVIEW 4 major objections 5 minor 1 cited by

Action Chunking with Transformers for Image-Based Spacecraft Guidance and Control

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read An imitation-learning policy trained on 100 expert demonstrations outperforms a meta-reinforcement-learning policy trained on 40 million interactions for image-based spacecraft docking, cutting mean distance error by about 29 percent and co

desk verdict Useful application of ACT to image-based spacecraft docking with a solid smoothness result, but the headline accuracy gain needs better statistics and a defined demonstration-selection rule before I'd trust it. read the letter →

arxiv 2509.04628 v1 pith:67R4TOYE submitted 2025-09-04 cs.RO cs.AI

Alejandro Posadas-Nava , Andrea Scorsoglio , Luca Ghilardi , Roberto Furfaro , Richard Linares This is my paper

classification cs.ROcs.AI

keywords imitationlearningactionchunkingwithtransformersspacecraftguidanceandcontrolimage-baseddockingmeta-reinforcementsampleefficiencytemporalensembling

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that a transformer-based imitation learner, trained on just 100 expert demonstrations of an image-based docking maneuver, outperforms a meta-reinforcement-learning agent trained on 40 million interactions. On the simulated ISS docking task, the Action Chunking with Transformers (ACT) policy achieves about a 28–29% lower mean terminal distance error and produces control commands that are 82% smoother by an L2 successive-action metric. The authors interpret this as evidence that a hybrid pipeline—meta-RL generates expert trajectories, imitation learning distills them—can yield deployable guidance policies with far less data and less control chatter than pure RL. The paper's own limitation section concedes that the controller still does not consistently meet the 10 cm accuracy threshold required for actual ISS docking.

What carries the argument

Action Chunking with Transformers (ACT): a transformer encoder-decoder paired with a conditional variational autoencoder. The encoder fuses ResNet-processed camera images, a linearly embedded spacecraft state vector, and a latent style variable z; the decoder predicts a chunk of k future actions rather than a single action. At inference, overlapping chunk predictions are combined with an exponentially weighted moving average (temporal ensembling), which the paper credits for smooth control and for mitigating compounding behavioral-cloning error.

What would settle it

Retrain the ACT policy from 100 rollouts drawn uniformly at random from the meta-RL baseline, with no success filtering, and compare mean terminal distance and smoothness on the same 100-episode evaluation; if the gap to RL shrinks or reverses, the claimed advantage rests on demonstration selection rather than action chunking.

Watch

Extended reading notes

Core claim

The central claim is that ACT, a transformer encoder-decoder combined with a conditional variational autoencoder that predicts temporally extended action chunks and blends overlapping chunks via temporal ensembling, can map camera images and state vectors to thrust and torque commands. Trained on 100 demonstrations generated by a meta-RL agent in the same Blender-based docking environment, the ACT policy beats the meta-RL baseline: mean terminal distance 0.992 m versus 1.391 m, best run 18 cm versus 87 cm, 43% versus 0% of runs within 80 cm, and a mean smoothness score of 1.71 versus 9.39. The authors attribute the gains to action chunking reducing the effective control horizon and to tempor

Load-bearing premise

The comparison assumes the 100 demonstrations are a fair sample of the meta-RL expert, but the paper only says 'the most successful' trajectories would be selected and never specifies the selection criterion; if demonstrations were filtered by final distance, the accuracy comparison is biased in ACT's favor.

Editorial extensions

If this is right

If the result holds, mission designers can obtain image-based docking controllers from a few hundred demonstrations, sidestepping the tens-of-millions-of-interactions data budgets typical of reinforcement learning.
The 82% reduction in control chatter means lower actuator wear and more stable commands, which matters for safety in close-proximity operations.
The hybrid loop—meta-RL generates demonstrations, imitation learning distills them, and the distilled policy can warm-start meta-RL—suggests the two paradigms are complementary for spacecraft GNC.
The controller's partial generalization to randomized initial conditions (37% of runs within 4 m) positions it as a viable coarse approach policy, though not yet a full docking solution at the 10 cm standard.
The method's computational lightness, asserted by the authors, supports the possibility of on-board deployment for image-based proximity operations.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial inference: the unspecified 'most successful' demonstration-selection rule is the main confound; a fair comparison would apply the same selection criterion to both the ACT training set and the RL evaluation set.
Editorial inference: the method's practical value may be strongest as a data-efficient warm start or coarse approach controller, with a separate fine-docking stage for the final meters, given the admitted failure to reach the 10 cm threshold.
Editorial inference: the same distillation recipe is testable on other image-based proximity operations, such as satellite servicing or asteroid rendezvous, where expert demonstrations are scarce and smooth commands are safety-critical.
Editorial inference: a direct ablation varying chunk size and temporal-ensemble decay would separate how much of the smoothness gain comes from chunking versus ensembling; the paper lists this as future work but does not test it.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short take: this is a reasonable application paper with one very solid statistical result and one under-supported headline claim. If you work on IL for spacecraft GNC, it's worth a read, but don't quote the 28-29% number without checking the details.

What's genuinely new here is using ACT to distill meta-RL demonstrations into an image-based 6-DOF docking controller. That's not in the literature as far as I know, though the authors' own reference [10] (Quinn et al., end-to-end IL for asteroid proximity ops) seems to undercut their claim that no one has done end-to-end IL with visual navigation before. The smoothness comparison is the best part: they use a per-episode L2 norm of successive action differences, run Welch's t-test after checking normality and heteroscedasticity in Appendix A, and report a large, significant effect. That part is methodologically sound.

The accuracy claim has real problems. Table 3 gives no confidence intervals; ACT's mean distance is better (0.992 vs 1.391 m) but its 95th and 99th percentiles are worse (1.763 vs 1.738, and 2.248 vs 1.810). The mean is driven by a few very good runs, and the two worst ACT runs are acknowledged but not handled. More fundamentally, the demonstrations are defined only as 'the most successful' trajectories of the meta-RL agent, with no selection criterion. The stress-test note is on target: if that means filtering by final distance, ACT is trained on a top-truncated slice of the RL distribution while the baseline is scored on the full distribution. That's an apples-to-oranges comparison. It's fixable: report the selection rule, evaluate the RL baseline on the same filtered subset, and give error bars or a bootstrap test. There are also numeric inconsistencies in the interaction count (6,300 vs 6,350 vs 6,530 in different places, and the conclusion says '40 environment interactions' instead of '40 million').

The limitations section is honest about not meeting the 10 cm docking threshold, which is good. No code or checkpoints are provided, so independent replication isn't possible from the preprint.

Who it's for: people working on imitation learning for spacecraft GNC, or anyone interested in how to evaluate learned controllers fairly. It deserves a serious referee, but the referee should ask for the missing analysis before accepting. I'd be comfortable citing it as an example of ACT applied to spacecraft docking once the numbers are cleaned up; not before. Send it to peer review, conditional on the authors addressing the demonstration-selection rule and the statistical gaps.

Referee Report

4 major / 5 minor

Summary. The paper proposes using Action Chunking with Transformers (ACT), an imitation-learning method, to learn a 6-DOF spacecraft docking policy from 100 expert demonstrations produced by a meta-RL agent. The method is evaluated in a simulated ISS docking environment with image and state observations. The authors claim that ACT trained on roughly 6,300 environment interactions outperforms a meta-RL baseline trained on 40 million interactions, improving mean final distance error by about 28–29% and producing 82% smoother control actions. The paper includes an architecture description, evaluation tables, trajectory figures, and a statistical analysis of control smoothness in Appendix A.

Significance. If the central claims hold, the paper would demonstrate an unusually large sample-efficiency gain for an image-based 6-DOF spacecraft GNC task, which would be of practical interest for onboard autonomy. A strength is that the smoothness comparison is properly tested with a Welch t-test, including normality and variance checks in Appendix A. The use of the Federici et al. environment also allows direct comparison with prior work. However, the accuracy claim is currently not supported at the same standard: the demonstration-selection rule is not specified, and Table 3 reports no confidence intervals or significance test for the distance improvement. The numerical inconsistencies regarding interaction counts also need correction before the quantitative claims can be accepted.

major comments (4)

[Introduction / Section IV] The demonstration-selection rule is unspecified. The Introduction states that 'only the most successful would be selected as demonstrations for IL training,' but no criterion—such as a final-distance threshold, reward quantile, or manual selection—is given anywhere in Section IV or the appendices. If demonstrations are filtered by task performance while the RL baseline is scored on all 100 evaluation rollouts, the reported 28–29% mean distance improvement could be an artifact of comparing a top-truncated training set against the full RL distribution. The authors must specify the exact selection criterion, report statistics of the demonstration pool versus the full meta-RL rollout distribution, and ensure that the ACT evaluation uses the same initial-state distribution as the RL evaluation.
[Table 3] The central accuracy claim lacks statistical support. Table 3 reports only means and percentiles for final distance and velocity, with no confidence intervals, standard errors, or hypothesis test. The text acknowledges that two ACT runs (2.50 m and 2.24 m) skew the mean. Moreover, the percentiles show that ACT is not uniformly better: at the 95th percentile ACT is worse than RL (1.763 m vs 1.738 m) and at the 99th percentile it is substantially worse (2.248 m vs 1.810 m). The authors should report bootstrap or analytic confidence intervals for the mean difference, a robust central measure such as the median, and a formal test (e.g., Mann–Whitney or bootstrap) for the distance comparison.
[In-Orbit Docking / Evaluation protocol] The evaluation protocol is ambiguous. The text says 'initial conditions were set to match the initial conditions seen in training,' but it is not clear whether this refers to the meta-RL training distribution, the ACT demonstration distribution, or a single fixed initial state. Appendix A refers to 'an identical initial state' across episodes, but Table 2 gives ranges (X in [-1,1], Y in [-26,-24], Z in [-1,1]), which suggests sampling within a box. The difference matters for interpreting the accuracy comparison and the generalization claim. The random-initialization results (Figure 3b) are also reported only for ACT, with no RL baseline, so they do not support a comparative generalization claim. The authors need to state exactly how initial states were sampled in each evaluation set and, ideally, compare ACT with RL under the same randomized protocol.
[Abstract / Table 1 / Introduction / Conclusion] The interaction counts are inconsistent across the manuscript. The abstract says 6,300 interactions for ACT; Table 1 says 6,350; the conclusion says 6,530. For the RL baseline, Table 1 reports 40,000,000 interactions, but the Introduction says 'approximately 610 k' and the conclusion contains '40environment interactions' (likely a typo for 40 million). These discrepancies directly affect the sample-efficiency claim and must be harmonized.

minor comments (5)

[Conclusion] Typo: '40environment interactions' should be '40 million environment interactions.'
[Abstract and Section IV] The improvement percentage is given as 28% in the Introduction and 29% in the Conclusion. Please make these consistent or explain the difference (e.g., mean vs. median).
[Appendix B] Typo: 'Paritally observable' should be 'Partially observable.'
[Figure 6] The visitation-density heatmaps would be easier to interpret with a colorbar and a clear statement of what quantity is plotted (e.g., normalized counts per bin).
[Section IV] The random-initialization test is described as a '4 m cube' in the text, but Table 2 gives ranges of 5 m in each axis (X and Z from -2.5 to 2.5, Y from -27.5 to -22.5). Please reconcile the cube size with the stated bounds.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported comparisons are empirical and the smoothing result is an explicitly acknowledged design consequence, not a hidden derivation from its own inputs.

full rationale

The paper's central claims are empirical. ACT is trained on 100 demonstrations produced by a meta-RL expert (Federici et al. [6]) and evaluated in the same simulated docking environment as the baseline. No parameter is fitted to the evaluation set and then renamed as a prediction; the accuracy comparison in Table 3 is a direct measurement of two policies on the same protocol. The sample-efficiency figures are reported from training runs, not derived from the evaluation. The smoothness improvement is partly by construction, since Eq. (3) temporal ensembling is explicitly designed to blend overlapping action chunks, and the paper states: 'The smoothing effect was the result of action chunking, which reduces the horizon of the trajectory, and temporal ensembling.' This is an acknowledged design mechanism, not a hidden circular derivation. The self-citations to Federici et al. supply the environment and the baseline; because that prior work includes an optical-bench validation and the current paper reproduces the same environment, the citation functions as independent external evidence rather than a load-bearing unverified premise. The unspecified demonstration-selection rule ('only the most successful would be selected') is a potential validity threat if selection were based on the evaluation metric, but the paper provides no evidence that this occurred, and under the hard rule requiring a specific reduction, this concern does not constitute demonstrated circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The ledger shows the paper relies on a small number of method hyperparameters that are not disclosed, plus domain assumptions inherited from the authors' prior meta-RL work. No new physical entities are introduced.

free parameters (4)

action chunk size k = not reported
Length of the action chunk in Eqs. 1 and 3; controls horizon and smoothing, a hand-selected ACT hyperparameter.
temporal ensemble decay coefficient m = not reported
Exponential weight in Eq. 3; directly determines how much smoothing is applied, hand-picked.
latent style dimension dz = not reported
CVAE latent size in Eq. 2; part of ACT architecture, not disclosed.
demonstration selection threshold = not specified
The Introduction says 'only the most successful' trajectories are selected for training, but no metric or cutoff is defined; this choice affects the fairness of the RL comparison.

assumptions (4)

domain assumption The meta-RL agent from Federici et al. [6] produces expert-quality demonstrations suitable for IL training.
Used in the Methodology and Introduction; if the teacher is suboptimal or poorly calibrated, the IL result inherits its biases.
domain assumption Hill-Clohessy-Wiltshire relative motion and Euler rotational dynamics adequately model the ISS docking scenario.
Reused from [6], stated in the In-Orbit Docking section; standard approximation for short-range rendezvous.
domain assumption Blender-rendered images are a faithful proxy for the navigation camera's observations.
The Visual Environment section assumes the rendered Sun-lit scene is physically accurate enough for policy learning and evaluation.
standard math ACT architecture and training follow Zhao et al. [16].
The paper cites [16] for the method and does not re-derive the transformer or CVAE; correctness of the learning algorithm is inherited from prior work.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Action Chunking with Transformers for Image-Based Spacecraft Guidance and Control." pith.science (2026). https://pith.science/paper/67R4TOYE

@misc{pith2026250904628,
  author       = {Pith},
  title        = {Pith review of: Action Chunking with Transformers for Image-Based Spacecraft Guidance and Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/67R4TOYE}},
  note         = {Machine review of arXiv:2509.04628}
}

read the original abstract

We present an imitation learning approach for spacecraft guidance, navigation, and control(GNC) that achieves high performance from limited data. Using only 100 expert demonstrations, equivalent to 6,300 environment interactions, our method, which implements Action Chunking with Transformers (ACT), learns a control policy that maps visual and state observations to thrust and torque commands. ACT generates smoother, more consistent trajectories than a meta-reinforcement learning (meta-RL) baseline trained with 40 million interactions. We evaluate ACT on a rendezvous task: in-orbit docking with the International Space Station (ISS). We show that our approach achieves greater accuracy, smoother control, and greater sample efficiency.

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

DIPOLE: Fusing Vision and Geometry for Robust Visuomotor Generalization
cs.RO 2025-11 conditional novelty 5.0 of 10

Fusing RGB and point-cloud inputs with training-time modality dropout plus cross-attention makes a diffusion visuomotor policy markedly more robust to visual and spatial shifts than unimodal or naively fused baselines.

Reference graph

Works this paper leans on

20 extracted references · 15 canonical work pages · cited by 1 Pith paper

[10]

End-to-End Imitation Learning for Optimal Asteroid Proximity Opera- tions,

P . Quinn, G. Nehma, and M. Tiwari, “End-to-End Imitation Learning for Optimal Asteroid Proximity Opera- tions,” 2025

work page 2025
[1]

Autonomy for Space Robots: Past, Present, and Future,

I. A. D. Nesnas, L. M. Fesq, and R. A. V olpe, “Autonomy for Space Robots: Past, Present, and Future,”Current Robotics Reports, V ol. 2, Sept. 2021, pp. 251–263, 10.1007/s43154-021-00057-2

doi:10.1007/s43154-021-00057-2 2021
[2]

Guidance, Navigation, and Control,

NASA Small Spacecraft Systems Virtual Institute, “Guidance, Navigation, and Control,”https://www. nasa.gov/smallsat-institute/sst-soa/guidance-navigation-and-control/, Apr

work page
[3]

Sparse Approximate Hamilton-Jacobi Solutions for Optimal Feedback Control with Terminal Constraints,

A. Jain, R. Eapen, and P . Singla, “Sparse Approximate Hamilton-Jacobi Solutions for Optimal Feedback Control with Terminal Constraints,”Proceedings of the 2023 62nd IEEE Conference on Decision and Control (CDC), 2023, pp. 1269–1274, 10.1109/CDC49753.2023.10384267

arXiv 2023
[4]

Franquiz,Attitude Determination & Control System Design and Implementation for a 6U CubeSat Proximity Operations Mission

F. Franquiz,Attitude Determination & Control System Design and Implementation for a 6U CubeSat Proximity Operations Mission. Doctoral dissertations and master’s theses, Embry-Riddle Aeronautical University, May

work page
[5]

Towards Resilient Tracking in Autonomous V ehicles: A Distributionally Robust Input and State Estimation Approach,

K. Azizi, K. Anurag, and W. Wan, “Towards Resilient Tracking in Autonomous V ehicles: A Distributionally Robust Input and State Estimation Approach,” 2025

work page 2025
[6]

Optical Bench V alidation of Meta-Reinforcement Learning for Image-Based Spacecraft Guidance and Navigation,

L. Federici, A. Scorsoglio, and R. Furfaro, “Optical Bench V alidation of Meta-Reinforcement Learning for Image-Based Spacecraft Guidance and Navigation,”Preprint, 2025, 10.22541/au.174253712.25691501/v1

work page arXiv 2025
[7]

IMAGE-BASED OPTIMAL POWERED DESCENT GUIDANCE VIA DEEP RECURRENT IMITA TION LEARNING,

L. Ghilardi, A. D’ambrosio, A. Scorsoglio, R. Furfaro, R. Linares, and F. Curti, “IMAGE-BASED OPTIMAL POWERED DESCENT GUIDANCE VIA DEEP RECURRENT IMITA TION LEARNING,”ASTRODYNAM- ICS 2020, Advances in the Astronautical Sciences(R. S. Wilson, J. Shan, K. C. Howell, and F. R. Hoots, eds.), 2021, pp. 2691–2706. AAS/AIAA Astrodynamics Specialist Conference, 2020

work page 2020

Show all 20 references

[8]

A Survey of Meta-Reinforcement Learning,

J. Beck, R. Vuorio, E. Z. Liu, Z. Xiong, L. Zintgraf, C. Finn, and S. Whiteson, “A Survey of Meta-Reinforcement Learning,”arXiv preprint, Aug. 2024. arXiv:2301.08028, 10.48550/arXiv.2301.08028

2024 doi arXiv
[9]

Autonomous Six-Degree-of-Freedom Spacecraft Docking Maneuvers via Reinforcement Learning,

C. E. Oestreich, R. Linares, and R. Gondhalekar, “Autonomous Six-Degree-of-Freedom Spacecraft Docking Maneuvers via Reinforcement Learning,” 2020

2020
[11]

Imitation Learning for Satellite Attitude Control under Unknown Perturbations,

Z. Zhang, H. Peng, and X. Bai, “Imitation Learning for Satellite Attitude Control under Unknown Perturbations,” 2025

2025
[12]

Learning Representations by Back-Propagating Errors,

D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “Learning Representations by Back-Propagating Errors,” Proceedings of the Neural Information Processing Systems (NeurIPS), 1988, pp. 318–326

1988
[13]

A Survey of Imitation Learning: Algorithms, Recent Developments, and Challenges,

M. Zare, P . M. Kebria, A. Khosravi, and S. Nahavandi, “A Survey of Imitation Learning: Algorithms, Recent Developments, and Challenges,” 2023

2023
[14]

Efﬁcient Reductions for Imitation Learning,

S. Ross and D. Bagnell, “Efﬁcient Reductions for Imitation Learning,”Proceedings of the Thirteenth Inter- national Conference on Articial Intelligence and Statistics(Y . W. Teh and M. Titterington, eds.), V ol. 9 of Proceedings of Machine Learning Research, Chia Laguna Resort...

2010
[15]

Grasping with Chopsticks: Combating Covariate Shift in Model-free Imitation Learning for Fine Manipulation,

L. Ke, J. Wang, T. Bhattacharjee, B. Boots, and S. Srinivasa, “Grasping with Chopsticks: Combating Covariate Shift in Model-free Imitation Learning for Fine Manipulation,” 2020

2020
[16]

Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware,

T. Z. Zhao, V . Kumar, S. Levine, and C. Finn, “Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware,”arXiv preprint, 2023. arXiv:2304.13705, 10.48550/arXiv.2304.13705

2023 doi arXiv
[17]

All Y ou Need Is Supervised Learning: From Imitation Learning to Meta-RL With Upside Down RL,

K. Arulkumaran, D. R. Ashley, J. Schmidhuber, and R. K. Srivastava, “All Y ou Need Is Supervised Learning: From Imitation Learning to Meta-RL With Upside Down RL,”arXiv preprint, Feb. 2022. arXiv:2202.11960, 10.48550/arXiv.2202.11960

2022 doi arXiv
[18]

BERT: Pre-Training of Deep Bidirectional Transformers for Language Understanding,

J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “BERT: Pre-Training of Deep Bidirectional Transformers for Language Understanding,”arXiv preprint, May 2019. arXiv:1810.04805, 10.48550/arXiv.1810.04805

2019 doi arXiv
[19]

International Docking System Standard (IDSS) Interface Deﬁnition Document,

NASA, “International Docking System Standard (IDSS) Interface Deﬁnition Document,” tech. rep., International Space Station Multilateral Coordination Board (MCB), Houston, TX, July 2022. Available athttps://www. internationaldockingstandard.com/. 10 APPENDIX A Smoothness metric...

2022
[2015]

Accessed: 2025-04-17

2025

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [10]

End-to-End Imitation Learning for Optimal Asteroid Proximity Opera- tions,

P . Quinn, G. Nehma, and M. Tiwari, “End-to-End Imitation Learning for Optimal Asteroid Proximity Opera- tions,” 2025

work page 2025

[2] [1]

Autonomy for Space Robots: Past, Present, and Future,

I. A. D. Nesnas, L. M. Fesq, and R. A. V olpe, “Autonomy for Space Robots: Past, Present, and Future,”Current Robotics Reports, V ol. 2, Sept. 2021, pp. 251–263, 10.1007/s43154-021-00057-2

doi:10.1007/s43154-021-00057-2 2021

[3] [2]

Guidance, Navigation, and Control,

NASA Small Spacecraft Systems Virtual Institute, “Guidance, Navigation, and Control,”https://www. nasa.gov/smallsat-institute/sst-soa/guidance-navigation-and-control/, Apr

work page

[4] [3]

Sparse Approximate Hamilton-Jacobi Solutions for Optimal Feedback Control with Terminal Constraints,

A. Jain, R. Eapen, and P . Singla, “Sparse Approximate Hamilton-Jacobi Solutions for Optimal Feedback Control with Terminal Constraints,”Proceedings of the 2023 62nd IEEE Conference on Decision and Control (CDC), 2023, pp. 1269–1274, 10.1109/CDC49753.2023.10384267

arXiv 2023

[5] [4]

Franquiz,Attitude Determination & Control System Design and Implementation for a 6U CubeSat Proximity Operations Mission

F. Franquiz,Attitude Determination & Control System Design and Implementation for a 6U CubeSat Proximity Operations Mission. Doctoral dissertations and master’s theses, Embry-Riddle Aeronautical University, May

work page

[6] [5]

Towards Resilient Tracking in Autonomous V ehicles: A Distributionally Robust Input and State Estimation Approach,

K. Azizi, K. Anurag, and W. Wan, “Towards Resilient Tracking in Autonomous V ehicles: A Distributionally Robust Input and State Estimation Approach,” 2025

work page 2025

[7] [6]

Optical Bench V alidation of Meta-Reinforcement Learning for Image-Based Spacecraft Guidance and Navigation,

L. Federici, A. Scorsoglio, and R. Furfaro, “Optical Bench V alidation of Meta-Reinforcement Learning for Image-Based Spacecraft Guidance and Navigation,”Preprint, 2025, 10.22541/au.174253712.25691501/v1

work page arXiv 2025

[8] [7]

IMAGE-BASED OPTIMAL POWERED DESCENT GUIDANCE VIA DEEP RECURRENT IMITA TION LEARNING,

L. Ghilardi, A. D’ambrosio, A. Scorsoglio, R. Furfaro, R. Linares, and F. Curti, “IMAGE-BASED OPTIMAL POWERED DESCENT GUIDANCE VIA DEEP RECURRENT IMITA TION LEARNING,”ASTRODYNAM- ICS 2020, Advances in the Astronautical Sciences(R. S. Wilson, J. Shan, K. C. Howell, and F. R. Hoots, eds.), 2021, pp. 2691–2706. AAS/AIAA Astrodynamics Specialist Conference, 2020

work page 2020