poster · WHERE 2026, Trento

Proprioceptive quadruped locomotion without gait priors, stage one: a learnable representation shaped by an auxiliary dynamics objective structures the locomotion manifold

Three stages towards velocity-adaptive locomotion under interoceptive rewards: learnable feature extraction, learnable discovery of motion primitives and the emergence of behaviour.

Janosch Bajorath · Malte Schilling

This page accompanies the poster and is being extended during and after the summer school (21–27 September 2026). Sections marked as such are still to come.

The road

Build it → understand it → take it apart → let it emerge. Today sits between the first two stations: a controller built from known ingredients that walks, trots and bounds, whose latent can be read.

Architecture and the three gradient sources

Diagram and description to come. An encoder over the proprioceptive input window produces the latent $z_t$; the base policy turns $z_t$, the command and PD gains into actions. Three gradient sources shape the encoder: reinforcement learning alone, plus a dynamics-reconstruction objective, plus a contact head.

Adaptivity: where the controller stands

At this stage, the environment schedules the kind of gait and its timing. The controller realises each pattern and crosses the switches. The wider ladder, from prescribed to self-owned gaits:

1. Randomised — the gait is a command: kind, frequency and duty are drawn at random; the controller learns to execute all of them, with no notion of which gait fits which demand.
2. Scheduled — today — the environment prescribes gait kind and timing; the controller realises each pattern and crosses the switches.
3. Training-only specification — the deployed controller receives no gait signal; the specification reaches the encoder only through decoder targets and the reward during training.
4. Self-specified — shaping rewards remain, but the specification they score against comes from the controller itself, not the environment.
5. Self-owned — the controller prescribes the gait from its dynamics and the command, and uncovers the gaits itself; no gait reward.

What the latent carries

Latent map with per-gait cycles and the switching segment — figures and animations to come.

How these were run

To come: the matched setup behind every figure — same task, rewards, commands and schedule; a shared graph; what varied was the gradient source.

Outlook

Next to remove: the environment clock, then the shaping reward. The policy owns its phase and frequency; gaits emerge from the embodiment; transitions from velocity alone.