Skills as reusable behaviours
We have now seen options (named skills) and goals (proposed destinations). Skill discovery asks the most ambitious version of the question: can an agent invent a useful repertoire of behaviours before it is ever told the task — purely by interacting with the world? If so, a later high-level controller can compose those pre-built skills into a hierarchy almost for free.
Diagram contrasting the three learning paradigms: supervised, unsupervised, and reinforcement learning.
Diversity is all you need (DIAYN)
The signature method is Diversity Is All You Need (DIAYN). Its insight is delightfully simple: a good skill is one you can tell apart from the others by looking at where it goes. So instead of a task reward, give the agent an intrinsic reward for making its skills distinguishable.
DIAYN's objective: maximize the mutual information between the skill code z and the states it visits, so different skills become easy to tell apart.
- Sample a latent skill code z at the start of an episode and condition the policy on it (much like a goal in goal-conditioned RL).
- Train a discriminator that tries to guess which z produced each visited state.
- Reward the policy when the discriminator guesses correctly — i.e. reward each skill for visiting states that uniquely identify it. Maximising this mutual information spreads the skills out to cover the state space.
DIAYN's pseudo-reward: the policy is paid when the discriminator q_φ confidently recovers the skill code z from the visited state s.
Astonishingly, with no task reward at all, DIAYN-style agents learn to walk, hop, run, and flip — distinct locomotion gaits emerge purely from the pressure to be different. Those gaits are exactly the kind of reusable primitives a hierarchy wants.
From discovered skills to a working hierarchy
Discovery and control then snap together. Freeze the learned skill set, and train a high-level policy that selects skills (or skill codes z) to maximise the real task reward. The hierarchy is now two clean layers: a discovered low level that knows how to move, and a learned high level that decides which behaviour to deploy when — the same manager-worker split from the feudal world, but with the worker's skills bootstrapped reward-free.
Because the skills were learned once and are task-agnostic, the same repertoire transfers across many downstream tasks — pay the exploration cost up front, reuse forever. This is why skill discovery is often pitched as a direct cure for the subgoal-discovery bottleneck and for the option collapse we met with Option-Critic.
The frontier: what's still genuinely hard
- Useful vs. merely diverse. Diversity objectives reward being different, not being useful. Skills may scatter to cover trivial state dimensions and miss the behaviours a real task needs — aligning discovery with downstream value is open.
- Deeper than two levels. Almost everything in practice is two layers. Stable learning of three or more levels of temporal abstraction remains largely unsolved.
- Non-stationarity between levels. While the worker is still changing, the manager is aiming at a moving target; co-adapting both layers stably is delicate.
- When does hierarchy even help? On many benchmarks a well-tuned flat agent matches a hierarchical one. Characterising which problems truly need hierarchy is an active debate.
Where to go next
You now hold the two great families of hierarchy — options (guides 2–3) and goals (guide 4) — plus the discovery machinery (this guide) that feeds both. The natural next steps are exploration (skill discovery is really directed exploration in disguise), transfer and meta-RL (where reusable skills shine), and multi-agent settings where managers coordinate whole teams.