Implicit Behavioral Cloning (IBC)

Re-implemented Implicit Behavioral Cloning (Florence et al., CoRL 2021) and compared it against an MSE regression baseline on two multimodal control tasks: a Particle 2D toy task and the block push environment from the IBC paper. Model: an energy-based EBMMLP that scores (observation, action) pairs, in contrast to the ExplicitMLP baseline that maps observations directly to actions. Training: InfoNCE loss so the expert action has lower energy than sampled negatives, teaching the network an implicit conditional distribution over actions rather than a single mean. Inference: action selection by minimizing energy via a DFO + Langevin dynamics sampler, letting the policy pick a single valid mode instead of averaging over conflicting demonstrations. Result: on the multimodal Particle 2D task, MSE averages over modes and produces invalid actions, while IBC correctly commits to one mode. On block push, closed-loop rollouts of the IBC policy track the expert distribution far more faithfully than MSE, and a DAgger variant further improves robustness.