Introduction
Multimodal affordance grounding is a critical capability for robotic agents, enabling them to identify functional regions of objects—such as the handle of a hammer or the blade of a knife—based on visual, structural, and semantic cues. While depth and language provide essential context beyond simple RGB imagery, existing multimodal models often rely on separate, heavy encoders for each data type. This "brute-force" fusion approach results in massive parameter counts and high computational overhead, hindering real-time deployment on resource-constrained robotic hardware.
The BiT-Align Framework
To address these efficiency challenges, we introduce BiT-Align (Bypass and Text-guided Attention for Multimodal Affordance Grounding). BiT-Align is a weakly-supervised framework that leverages two novel mechanisms to achieve state-of-the-art performance with minimal parameters.
Bypass Prompt Module (BPM)
Instead of a dedicated depth encoder, BiT-Align utilizes the Bypass Prompt Module. BPM treats auxiliary depth data as a "prompt" for the primary RGB Vision Transformer (DINO-ViT-S). Geometric features from depth maps are progressively injected into the intermediate layers of the RGB encoder, modulating visual features without the need for a parallel heavy backbone.
Text Feature Guidance (TFG)
The Text Feature Guidance module uses semantic embeddings (via CLIP) to dynamically control the model's focus. By matching text features (e.g., "hold" or "cut with") to specific attention head masks within the image encoder, the model suppresses irrelevant background noise and highlights regions pertinent to the specific task.
Experimental Setup
The framework was evaluated on two major benchmarks: AGD20K and HICO-IIF. Since these datasets lack native depth information, high-quality pseudo-depth maps were generated using Depth Anything V2. The model learns from exocentric (third-person) interaction images and successfully transfers this knowledge to egocentric (first-person) robotic views.
Key Results and Discussion
BiT-Align demonstrates that multimodal learning does not require excessive parameter bloat. The model achieved state-of-the-art accuracy, improving the Kullback-Leibler Divergence (KLD) metric on AGD20K by 6.0% compared to previous baselines.
The most significant contribution is the model's efficiency. BiT-Align requires only 7.7M parameters, representing an 88.8% reduction over the previous state-of-the-art. Computational complexity (GFLOPs) was reduced by 72.4%, and inference time dropped to 10.0 ms per image. Real-world validation on a UR5 robotic arm confirmed the model's ability to guide functional grasping and tool use in real-time.
Conclusion
BiT-Align proves that treating auxiliary modalities as prompts and using semantic guidance for attention can drastically reduce the resource requirements of multimodal models. While future work will address sensitivity to depth noise and explore larger backbones, BiT-Align represents a major step toward practical, real-time affordance grounding for robotics.
Summary updated