3D-VCD：无需重训练即可消除3D具身智能体幻觉的突破性方法

This post introduces 3D-VCD (3D Visual Contrastive Decoding), the first hallucination elimination framework for 3D embodied agent reasoning. It addresses the critical problem of hallucinations (descriptions/decisions inconsistent with real 3D environments) in key applications like robot navigation and autonomous driving. The core advantage is that it works at inference time without retraining the base model, significantly improving grounded reasoning performance on benchmarks like 3D-POPE and HEAL.

When LLMs become the 'brain' of 3D embodied agents, hallucinations emerge as a severe issue. Unlike 2D (object misidentification/attribute errors), 3D hallucinations include:

• Object existence: Claiming non-existent objects
• Spatial layout: Wrong relative positions of objects
• Geometric grounding: Incorrect size/shape/orientation of objects Existing 2D methods rely on pixel-level perturbations which are ineffective for structured 3D inputs (scene graphs/point clouds). 3D hallucinations often stem from over-reliance on language priors rather than actual 3D observations.

3D-VCD's key innovation is building semantically and geometrically perturbed 'distorted' 3D scenes using two strategies:

1. Semantic perturbation: Randomly replace object category labels (e.g., chair → table) to test if the model uses visual evidence or language priors.
2. Geometric perturbation: Adjust object coordinates, bounding box size/shape, or orientation to test spatial relation understanding based on 3D geometry.

The contrastive decoding process involves three steps:

1. Parallel forward propagation: Model processes original and multiple distorted scenes to get token probability distributions.
2. Sensitivity analysis: Calculate probability differences of each token between original and distorted scenes. Tokens with little change are likely hallucinations (driven by language priors).
3. Dynamic suppression: Adjust token sampling probabilities—suppress insensitive tokens and enhance responsive ones to ensure consistency with real 3D environments.

3D-VCD was tested on two benchmarks:

• 3D-POPE: Reduced hallucination rate by over 30% while maintaining high answer coverage.
• HEAL: Improved complex reasoning accuracy by 15-20%. Ablation studies show combining semantic and geometric perturbations yields best results; moderate perturbation strength is key. It adds ~2-3x inference overhead but requires no retraining, making deployment easy.

3D-VCD contributes to the field by shifting from 2D pixel perturbations to 3D structured scene perturbations, offering a practical no-retraining solution. Applications include:

• Home service robots: Reduce navigation/object recognition errors.
• Autonomous driving: Enhance environmental perception reliability.
• AR: Improve virtual-reality fusion stability.
• Industrial maintenance: Reduce false/missed detections in equipment inspection.

Current limitations:

• Computational overhead (2-3x inference cost).
• Perturbations are object-level; need finer-grained strategies (textures, room partitions).
• Designed for static scenes; dynamic environments require time-consistent mechanisms.
• Need to extend to multi-modal inputs (visual + language + tactile feedback). Future work will address these to make 3D-VCD more efficient and adaptable.