A robot that needs to understand a scene in three dimensions typically runs several perception tasks at once, segmenting what things are while estimating how far away they are. Sharing one strong feature encoder across those tasks is now standard practice, and vision foundation models are the usual choice.

The authors locate the bottleneck downstream of that encoder. The decoding strategies that turn shared features back into per-pixel predictions have become the expensive part, with the Dense Prediction Transformer as the standard and costly option. A robot has a power budget and a latency budget, and neither is generous.

DPNeXt is their lighter replacement, a streamlined multi-scale fusion decoder built from dual depthwise separable inverted bottlenecks. It is an unfashionable place to optimise, since the encoder gets the attention, but on a machine that has to perceive continuously it is where the cost actually accumulates.