Two Minute Papers · 2026-08-02 · notable
Two Minute Papers — 'AI Learns Why Copying Humans Isn't Enough'
Two Minute Papers' latest episode walks through a paper on why pure imitation learning tops out — and what agents that go beyond behavioural cloning learn instead.
Károly Zsolnai-Fehér explains why AI systems that only copy human demonstrations plateau — and what to do about it.
What is it?
Two Minute Papers' newest video, 'AI Learns Why Copying Humans Isn't Enough,' unpacks a paper on the limits of imitation learning — the trick of training agents by cloning recorded human behaviour. The video runs Károly Zsolnai-Fehér's usual side-by-side clips of policies learned from demonstrations versus policies that go further.
How does it work?
Behavioural cloning turns human trajectories into supervised training data: given this state, output that action. The video walks through why that plateaus — off-distribution states, compounding error, missing recovery behaviours — and shows what happens when the agent adds its own feedback signal on top of the demonstrations.
Why does it matter?
Imitation learning is the default recipe for humanoid robots, self-driving stacks, and increasingly agentic assistants trained on human traces. This Two Minute Papers episode is a quick, visual entry point to the failure mode every team hits once demos alone stop scaling, and the paper it summarises points at where the next round of fixes is coming from.
Who is it for?
Robotics and RL practitioners, agent-training teams, ML students getting into imitation learning.