Option (b) feels smarter in the short term — it gives you immediate “proof” that your RGB-only model is making the right calls. The extra sensors act as ground truth during training, helping the model learn faster and fail more safely, even if it runs blind at inference.