The sensor modality is in the claim, and that is the first thing to note about GM's grant US11204417B2, "Selective attention mechanism for improved perception sensor performance in vehicular applications" (issued December 21, 2021). The CPC is radar-forward — G01S 13/867 and G01S 13/865 are radar-imaging and combined radar/optical classes, G01S 13/931 is automotive radar — fused with G06K 9/00791, road-scene recognition. So this is an attention layer applied to a named, radar-anchored perception system, not a generic algorithm.
'Selective attention' is borrowed from machine-learning vocabulary, and the enabling question is whether the claim does real work or merely invokes the buzzword. Reading the construct as an attention mechanism that allocates perception effort to salient scene regions, the answer leans enabling: the value is the focusing logic, and the radar classes ground it in a concrete sensing context. The independent claim's defensible core is the mechanism tying scene saliency to sensor or compute allocation.
Strategically, this is GM staking the same efficiency frontier as Zoox's detection-optimization work from earlier in 2021 — different mechanism, same problem. As perception models grew heavier, the industry hit a wall on uniformly processing every pixel and every return at frame rate. Attention is the standard answer, and getting a vehicular-application grant on it is a sensible defensive and offensive position in the perception-compute fight.
The teardown verdict: the breadth is in the independent claim, but the durable protection is in the dependents that tie attention to specific radar signal characteristics. Do not read 'attention mechanism' as a deployed feature — it is a claimed method. What the December 2021 date tells you is that by the end of that year, attention-based perception had matured from research idea to patented vehicular IP, and GM moved to hold a radar-anchored corner of it.