Setting: Instance-level 6D pose for CAD objects unseen during training, without pose-specific fine-tuning

Input: RGB scene + CAD-rendered templates

Summary: ZS6D extracts descriptors from queries and CAD-rendered templates with a pretrained Vision Transformer, matches template viewpoints, establishes local 2D–3D correspondences, then solves RANSAC-PnP. Requires no task-specific fine-tuning on LM-O, YCB-V, or T-LESS.

Setting: 6D pose and metric size of unseen within-category instances from one RGB image

Input: monocular RGB

Summary: DMSR uses a pretrained monocular geometry estimator for local 3D cues and reliable 2D–3D correspondences. An independent branch recovers metric scale from category statistics, followed by RANSAC-PnP, directly addressing RGB-only category-level scale ambiguity.

Setting: Real-time category-level pose / size through implicit 3D perception from monocular RGB

Input: monocular RGB

Summary: Feature Auxiliary Perception Network learns coarse 3D features through geometry supervision, then refines them with pose and size constraints. On REAL275, reports gains of 14.7 percentage points at 10°10cm and 11.4 percentage points at IoU50, maintaining real-time inference.

Setting: Reduces dense real 6D pose annotations and jointly learns initialization and refinement

Input: RGB / RGB-D object observation

Summary: Joint losses couple self-supervised pose representations and refinement so initialization and subsequent correction are learned together. Primarily reduces industrial-object 6D annotation requirements, rather than eliminating all object priors at test time.

Setting: Given an object NeRF, directly recovers camera–object 6DoF pose from an arbitrary query image

Input: single RGB + object NeRF

Summary: IFFNeRF samples surface points from NeRF through Metropolis-Hastings and projects rays, learns query-to-ray-bundle correspondences with attention, then solves pose by least squares. Reports 34 FPS and substantially lower angular and translation errors than iNeRF.

Setting: Rapid object-NeRF construction from a few multiview images, followed by real-time pose for new images

Input: single RGB + object NeRF

Summary: Renders RGB / depth from an initial viewpoint, matches query and rendering with LoFTR or similar methods, then lifts 2D–2D matches into 2D–3D correspondences using NeRF depth. 3D consistency mining and occlusion sampling filter false geometry. Reports approximately 90× speedup and 6 FPS.

Setting: A robot scans a scene from scratch and localizes multiple known tiny objects in the reconstructed density field

Input: wrist RGB video + known 3D models

Summary: A wrist-mounted monocular camera actively scans the scene, efficient Neural Graphics Primitives reconstruct a density field, and known object models fit directly to it. Automatically detects and estimates multiple objects, reporting approximately 1 mm accuracy for small reflective industrial parts.

Setting: Actively collects multiview observations during arm motion and performs manipulation without depth sensors

Input: eye-in-hand monocular RGB

Summary: RGBManip actively acquires multiple views with a gripper-mounted monocular camera and estimates object 6D pose from images. A reinforcement-learning policy jointly optimizes viewpoints, pose accuracy, and manipulation time, evaluating image-only manipulation in simulation and real systems.

Setting: Interpretable physical-world models identify transparent objects and estimate pose from weak sensory observations

Input: poor / partial sensor observations + symbolic physical knowledge

Summary: NaivPhys4RP uses cognitive simulation to construct possible worlds satisfying commonsense, causal, and physical constraints, rather than large training datasets. White-box reasoning explains observations and recovers transparent-object identity and 6D pose, emphasizing interpretable, safe perception beyond standard BOP ranking.

Setting: Simulation-only training for target-workpiece detection and 6D pose in real randomly arranged bins

Input: RGB-D / point cloud + CAD-based simulation

Summary: Physics simulation and target CAD automatically generate stacked-bin scenes, instance segmentations, and pose annotations. A localization system trains only on synthetic data and deploys to real random bin picking. Evaluation should consider real pose accuracy, domain gaps, and grasp success together.

Setting: Continuously aligns a known 3D object model directly to asynchronous events during fast camera motion

Input: event stream + object 3D model

Summary: Constructs a new event representation, predicts brightness increments through object-motion interpolation, and directly registers them to projected 3D models. Evaluates tracking duration and high-speed robustness, operating at camera speeds around 2 m/s.

Setting: Joint multi-object detection and 6D pose in dynamic, occluded bin videos

Input: RGB video sequence

Summary: Processes video sequences to detect all objects and estimate poses in one forward pass. Temporal fusion aggregates object embeddings and parameters across time, improving detection and pose on SynPick and YCB-Video.

Setting: Learns multiple valid inter-object relative poses for one task from 10–20 demonstrations

Input: paired object point clouds + demonstrations

Summary: Extends TAX-Pose geometric-relation reasoning with a spatially grounded distributional demonstration latent space for multimodal relative placements. With only 10–20 demonstrations and no manual semantic labels, generates precise task-related cross-poses across different within-category objects.

Setting: High-frequency online 6DoF tracking of fast known rigid objects with an event camera alone

Input: asynchronous events + object mesh

Summary: EDOPT projects object meshes into event-compatible templates, maintains future pose hypotheses, and adapts tracking updates to varying event rates. The paper claims the first event-only 6DoF object-pose method, with accuracy approaching deep networks combining frames, depth, and events.