Monocular 6-DoF pose estimation of non-cooperative targets is important for on-orbit servicing and debris removal. A single-image estimator can confuse near-symmetric spacecraft orientations, and tracking can preserve an incorrect pose. We present Hypothesis-Anchored Tracking (HAT), a causal framework that uses inter-frame motion to select among competing CAD-based pose hypotheses before alignment and fusion. Rather than independently choosing the highest-scoring hypothesis in each image, HAT retains competing orientation histories and selects a pose to anchor the relative trajectory estimated by monocular SLAM. Sparse anchors and pose fusion provide per-frame estimates after initialization without revising past outputs. The method requires only a calibrated RGB sequence, a metric CAD model, and target image regions, which can be supplied by detection or segmentation. The pretrained pose and SLAM networks require no target-specific training or fine-tuning. We evaluate two versions, Mega-HAT and Pico-HAT, using MegaPose and PicoPose, on SPARK-2024, SwissCube and SHIRT, with YCB-Video assessing performance outside the space domain. Using one temporal configuration per method, the arithmetic means of the four dataset-wise comparisons show 9.4% lower mean pose error and 3.76 times the sustained input FPS for Mega-HAT relative to independent MegaPose, and 23.9% lower mean pose error and 2.42 times the FPS for Pico-HAT relative to independent PicoPose. Mega-HAT ablations on SPARK and an offline reference examine component contributions and the effect of revising past estimates.
Figures & tables
Parameter group
Mega-HAT
Pico-HAT
Absolute anchors
gap 4, K=10 , five refiner iterations
gap 6, K=5 , PicoPose stages 1–3, no MegaPose refiner
Decoder
α=3 , μ=0.3
α=100 , μ=0.3
Fusion
w=0.002 , D=75∘ warm-up 4, Huber scale 0.3
w=0.1 , D=30∘ warm-up 4, Huber scale 0.3
TABLE I: Shared temporal configurations for the four-dataset benchmark. Each method uses the same settings across datasets.
Configuration
Et(m)
Eq∘
Ep
Mega-HAT , causal
0.442
10.82
0.2355
Mega-HAT , all stages offline
0.430
9.66
0.2125
Mega-HAT , offline + dense refinement
0.348
9.52
0.2016
Pico-HAT , causal, tuned
0.123
4.19
0.0886
Pico-HAT , offline, tuned
0.104
2.81
0.0618
Motion-aware ViT [ 12 ] , offline
0.137
4.69
0.0990
TABLE II: Causal and offline HAT on synthetic SPARK, with tuned temporal settings.
Fig. 2: Rotation error and signed depth error T^z−Tz of appearance-selected MegaPose hypotheses on SPARK-2024, after refinement and before temporal selection. Markers show medians in 2m bins of ground-truth depth, and shading shows interquartile ranges. Negative depth errors indicate underestimated depth.
Configuration
Et(m)
Eq∘
Ep
MegaPose tracking
0.471
28.13
0.5387
Keyframe anchors + SLAM fusion
0.490
18.77
0.3788
+ Viterbi basin selection (Stage 3)
0.469
12.20
0.2619
+ size-weighted unary
0.451
11.63
0.2504
+ disagreement gate ( D=75∘ )
0.442
10.82
0.2355
TABLE III: Cumulative component ablations of Mega-HAT on SPARK, causal throughout.
Method
Ep
SD
Et ( m )
Eq ( ∘ )
FPS
MegaPose (per frame)
0.413
0.808
0.469
21.0
0.82
PicoPose
0.098
0.344
0.112
4.8
2.74
GigaPose + ref.
0.410
0.811
0.333
21.5
1.66
RGBTrack
2.169
1.065
1.804
111.0
2.34
SRT3D + init.
0.650
0.943
1.311
29.1
131.61
Mega-HAT
0.325
0.648
0.463
15.9
2.27
TABLE IV: Means and sample SD summarize each dataset. Bold denotes the best result in each column per dataset (lowest errors and SD, highest FPS). Italics denote video baselines; our methods appear below the divider. “+ ref.” adds five MegaPose refinement iterations. FPS includes decoding after warm-up and excludes detector generation and loading. Mega-HAT uses the shared temporal configuration. Evaluation and timing protocols are specified in Sec. VI .
Fig. 3: Mean Ep (bars), one sample SD (upper whiskers), and sustained input FPS (diamonds). Our implementations are colored, with larger outlined diamonds. Baselines are gray. Numerical labels mark whiskers above 4 and rates above 20FPS . Throughput is measured on separate stream subsets. Table IV specifies labels and evaluation scope.
Monocular 6-DoF spacecraft pose estimation methods predominantly process individual frames, discarding the temporal information present in an image sequence acquired during spacecraft manoeuvres. Few temporal approaches require full backbone fine-tuning or auxiliary optical flow networks, risking catastrophic forgetting or increasing computational cost, respectively. We propose TALON (Token-Aligned Lightweight adapters for Orbital Navigation): spatiotemporal 3D adapters injected before the self-attention layers of a frozen ViT vision transformer, combined with a patch-token alignment loss that geometrically grounds the adapted features to keypoint structure through a prototype-conditioned KL-divergence objective. Pre-attention placement allows the frozen attention to reason over temporally enriched tokens, achieving stronger performance with a single adapter per block than post-attention alternatives. The alignment loss shapes the intermediate representations so that each keypoint induces a spatially precise activation in the token field, while the framework adds less than 5% parameters to the frozen backbone. On SPADES dataset, TALON reduces the pose error by 50% over the prior state-of-the-art, and on SwissCube dataset it surpasses the prior best by 21.8% in ADD-0.1d accuracy. Zero-shot cross-domain evaluation from sim-to-real on SPARK real data reduces pose error by 4.7x, and ablations characterise the role of adapter depth across in-domain and cross-domain settings.
6-DoF pose estimation is a critical task in autonomous rendezvous and proximity operations. In the case of an unknown target, this task becomes challenging as it shall be paired with the reconstruction of the target shape model. In this article, we propose a novel framework for single-shot shape and pose estimation of unknown spacecraft objects. Given a single image, we first reconstruct a 3D shape model of the target, then estimate the relative six-degrees-of-freedom pose by learning dense 2D-3D correspondences. The image features are extracted using a frozen DINOv3 vision transformer, while the geometric features are computed from the reconstructed point cloud using a trainable dynamic graph convolutional neural network encoder. A dual-stream transformer matcher refines descriptors through alternating self- and cross-attention, producing soft correspondences that are passed to a Perspective-n-Point solver for pose recovery. We evaluate the method on the SPE3R dataset and consider FoundationPose as a representative baseline for current state-of-the-art capabilities. Results show reliable pose estimates achieving 0.157 degrees mean pointing error using only a single image and reconstructed geometry, demonstrating strong generalization to unseen spacecraft.
Josiane Uwumukiza, Jocelyn Zhao, Giovanni Lavezzi +5
Undergraduate Student, Wellesley College, Wellesley, MA 02481, USA · Undergraduate Student, Massachusetts Institute of Technology, Cambridge, MA 02139, USA · Research Scientist, Massachusetts Institute of Technology, Cambridge, MA 02139, USA +5
We present an improvement on previous spacecraft pose estimation architectures that results in the lowest published mean rotation errors we know of on the SPEED+ lightbox and sunlamp test sets for a known, non-cooperative spacecraft. By using a previously established heatmap-based pose estimation architecture and adapting a large self-supervised ViT foundation model (DINOv3) in place of the smaller convolutional and ViT encoders of previous work, we show that pose estimation accuracy improves from 300M to 840M parameters with no saturation yet observed. We also evaluate our 840M model on a Jetson Orin NX 16GB, measuring single-pass network inference at 133.8 ms per crop with a board draw of 32.0 W. These measurements demonstrate embedded inference feasibility on a processor family with orbital flight heritage. Our resulting model outperforms previous models across lightbox and sunlamp domains while training only on synthetic data. Our best model, using DINOv3 840M adapted with LoRA as the encoder (rank 64, three-seed ensemble with four-rotation test-time augmentation), results in 1.56∘ mean rotation error on sunlamp and 1.17∘ on lightbox, compared to the previous best mean rotation errors we know of on these test sets, 2.66∘ and 1.75∘ by EagerNet.
John Church, Vazghen Nikolian
Both authors contributed equally to this work; author order is alphabetical.