为什么给机器人装上昂贵的触觉传感器，反而让它变笨了？

Core Insights - The article discusses a new approach to multi-modal robot learning that addresses the limitations of traditional feature concatenation methods, which often fail in tasks requiring tactile feedback [3][5][33] - The proposed solution involves using compositional policies, where each sensory modality is trained as a separate expert, allowing for more flexible and robust integration of sensory data [9][12][33] Limitations of Current Methods - Traditional multi-modal robot learning typically relies on feature concatenation, which combines all sensor embeddings into a single vector, leading to significant performance drops in tasks requiring tactile information [5][16] - The feature concatenation method treats rare but critical tactile signals as noise, resulting in a drastic decrease in success rates from 35% to 5% when tactile data is added [3][16] Proposed Solutions - The new approach involves training separate expert policies for each modality, allowing for independent learning and reducing interference between modalities [9][12] - This modular design enables easy addition or removal of sensors without the need to retrain the entire system, thus lowering retraining costs and enhancing system robustness [13][16] Performance Results - The proposed method achieved an average success rate of 66% across four RLBench simulation tasks, outperforming single-modal strategies (49%) and feature concatenation (56%) [29] - In specific tasks, the method demonstrated a success rate of 65% for occluded marker picking, compared to 35% for RGB-only and 5% for the concatenation method [34] Robustness and Adaptability - The system shows robustness to runtime disturbances, such as sudden object removal, and can adapt by reallocating weights to remaining functional sensors [21][23] - It maintains stable performance even when simulating sensor failures, demonstrating the effectiveness of the routing network in managing consensus weights [23][27]