End-to-end driving models can be merged after a single upload: FedE2E scores 55.85 on Bench2Drive closed-loop evaluation
Synopsis
Federated training for end-to-end driving has usually needed multiple communication rounds and open-loop error metrics, which do not capture error accumulation during continuous driving. FedE2E has each client upload its model weights once, then distills knowledge from those teacher models on server-side data through multi-client calibration, anchor-based alignment and token-wise consensus distillation. On the CARLA Bench2Drive closed-loop benchmark it reaches a Driving Score of 55.85 in the one-shot setting, 4.07 points above the distillation-based FedDF, while remaining below the centralized reference of 63.30.
Interpretation
An end-to-end driving model can now be federated with a single weight upload per participant, and the server no longer needs clients for later training. Earlier federated methods for end-to-end driving typically relied on multiple rounds of client-server communication, while existing one-shot methods mainly target classification and have seen limited exploration in end-to-end driving. Closed-loop evaluation on 220 routes of Bench2Drive in CARLA 0.9.15 gives FedE2E a one-shot Driving Score of 55.85, above the 51.78 of FedDF, which also uses knowledge distillation, and above FedAvg, FedRDN and FedProx.
The server treats several independently trained client models as teachers and distills their driving knowledge into a student model, the global model. Multi-client calibration first re-estimates each teacher's batch-normalization statistics on the server distillation data, anchor-based alignment then maps teacher features into the student space with orthogonal transforms, and token-wise consensus distillation weights and fuses teachers by their feature agreement. Experiments build a non-IID setting from four Bench2Drive towns, with CIL++ client models on a ResNet34 backbone, and evaluate five driving abilities: overtaking, merging, emergency brake, give way and traffic signs.
The global model from one-shot fusion can also initialize later federated training. Continuing with FedAvg after FedE2E fusion reaches a Driving Score of 60.35 after five communication rounds, above FedAvg at 58.54 and FedDF at 58.77. The per-round results show 59.58 at round three, already above the best five-round baseline result of 58.77, cutting the rounds needed for that level from five to three.
Perspective
Evaluation runs in the CARLA 0.9.15 simulator on 220 routes of the Bench2Drive benchmark, with training and testing on different towns and coverage of five driving abilities: overtaking, merging, emergency brake, give way and traffic signs; all client models use the CIL++ architecture.
In the one-shot setting FedE2E's Driving Score of 55.85 remains 7.45 points below the centralized 63.30; FedDF is better on give way and FedRDN scores highest on traffic signs; the choice of server-side distillation dataset affects results, with Driving Scores ranging from 54.03 to 57.47 across the four settings.
