First steps — learning to walk

So far, LifeLabs has shown shapes that persist (Lenia) and behaviors that emerge from a population (the evolutionary agents). Today there is a single creature, a single body, a single life — and a more intimate question: can you learn to walk when no one ever wrote down how? This is milestone 4, the last of the inaugural series, and it carries the name that stirs the most dreams and the most fantasies: the avatar.

A body, an empty brain, no gait written anywhere

The avatar is a bipedal body simulated in a physics engine (MuJoCo, the Walker2d environment): a torso and two articulated legs of thigh–shin–foot. Nothing more. Its “brain” is a neural network that, at the first instant, contains only noise: it sends each joint random commands. We coded no walking trajectory, no leg cycle, no notion of balance. We only defined a reward — moving forward earns points, falling costs them, staying alive pays a little — and let the agent try, fail, and try again, hundreds of thousands of times.

That is reinforcement learning (here, the PPO algorithm). Unlike the evolutionary agents, there is no population being ranked and mutated from one generation to the next: there is a single individual who, through its own experience, tunes the weights of its network to harvest more reward. And for once the word “learn” is used in its exact technical sense: the network changes its own parameters in order to improve a score. Nothing more, nothing less. No consciousness, no intention in there — a set of numbers adjusting itself so that a body falls less and travels more.

What this run shows

The starting point pulls no punches. At 8,000 training steps, mean reward is negative (−1.0), episodes last barely 18 steps before the fall, and net displacement is negative (−0.15): the avatar convulses, thrashes its legs, and collapses almost at once. A purely random policy, for comparison, gets nowhere either (distance −0.13).

Then, over 500,000 steps, something is built — and the seven checkpoint videos are the film of it, in order:

• Checkpoint 1 (~74k steps): the avatar first learned to stop falling immediately. Episodes stretch to ~165 steps and reward climbs to ~144. But net displacement stays near zero (−0.16): it stands, it sways, it goes nowhere yet. Learning to walk begins with learning not to topple.
• Checkpoints 2 to 4 (~147k → ~295k steps): walking begins. Displacement turns positive (+0.77, then +1.38, then +1.57) — the avatar moves clearly forward. The motion is still jerky, but the direction is there.
• Checkpoints 5 to 7 (~369k → ~516k steps): locomotion consolidates. Distance +2.09, then +2.53, finally +2.79; episodes reach 240 steps. The avatar genuinely walks forward, in a sustained way.

In final numbers: reward went from −1 to 587, episode length from 18 to 240 steps (thirteen times longer on its feet), and distance from −0.15 to +2.79 — from an instant fall to a held forward walk. This is a real success, and it deserves to be stated without embellishment: the avatar learned to balance and then move forward, but it is not yet a perfect or endless gait. It walks for a limited stretch, then eventually loses its balance — 240 steps, not the 1,000 available. For a long, regular stride you would need to push training to several million steps; 500,000 is enough to see walking be born, not to polish it. That is exactly the intended dial, and it is documented.

The most honest thing is to watch. The final model’s gait replays in 3D, right in your browser, on the experiment page — spin the camera around the avatar as it moves. The seven checkpoint videos are there too, side by side: the progression spasms → balance → walk reads at a glance. All the curves are on the full run page.

What it is not — and how it differs

We should name precisely what we’re watching. Unlike the evolutionary agents, nothing here emerges from a population: no selection, no mutation, no generations. A single individual learns over its own “lifetime” by adjusting its network from the reward it receives. That is the difference between evolution (Darwin sorts a population across many generations) and learning (one brain tunes itself over a single existence).

And unlike Lenia, this is not a shape emerging from a fixed local rule. In Lenia the rule never changes and a creature appears from it; here it is the rule itself — the network’s weights — that changes constantly, because the rule is what’s being learned. So it isn’t, strictly speaking, emergence: it is learning. The complexity comes neither from the rule (Lenia) nor from selection (agents), but from the repeated experience of one body.

None of this says anything about the avatar’s consciousness. A network minimizing a cost and harvesting a reward has no known subjective experience. What is unsettling is not that it “wants” to walk — it wants nothing — but that walking, that gesture we think so natural, can arise from a heap of numbers that does nothing but try and correct.

The end of a beginning

This post closes milestone 4 and, with it, the first series of LifeLabs experiments: a form that sustains itself (Lenia), a behavior that evolves (the agents), a skill that is learned (the avatar). Three very different ways of bringing something life-like into being without ever programming it directly. Everything is reproducible from seed 42, and the --timesteps dial is still there, ready to turn these hesitant first steps into a real stride for anyone willing to spend the compute hours.