DITTO-style latent divergence: researchers debate imitation learning for autonomous driving
BrantonDeMoss · x · 2026-10-24
A technical exchange in the self-play driving research community: asked why no explicit reward function is used, Yassine explained that basic rules (collisions, lane departures) can't capture good driving — traffic lights, signs, exits would all need modeling, so the team currently relies on a goal of the ego's future state.
Branton DeMoss proposed a fix: measure divergence against expert trajectory latents in a world model's latent space, turning imitation into on-policy RL. He cited his paper DITTO (arXiv:2302.03086), which trains a world model, unrolls the agent from expert start states, and penalizes multi-step latent divergence from expert data — provably inducing imitation learning and achieving SOTA performance and sample efficiency on Atari from pixels without online environment access.
More from Research
- RLCR trains LLM agents as organizations: emergent hierarchies and division of labour — algo_diver · 2026-10-24
- Simplified Matching Distillation: few-step diffusion without adversarial training, validated up to 33B — ArashVahdat · 2026-10-24
- tszzl predicts stunning biology breakthroughs within years, cancer included — tszzl · 2026-10-24
- rei_labs introduces CLP, a coupled learning-and-planning method for data-scarce training — burny_tech · 2026-10-24
- Palmer Luckey: a parrot's tiny brain shows efficiency, not scale, may unlock AI's next leap — NaveenGRao · 2026-10-24
- Freeform Preference Learning Lets Robot Annotators Compare Trajectories Along Natural-Language Axes — micoolcho · 2026-10-24