The playground opens with three choices: drive Asimov, assign scripted robot routines, or watch the Microducks. Nothing starts until you choose. Larger, labelled touch controls make it easier to hold a direction or slide between forward, back and turning.
Asimov keeps the latest walking and recovery policies. The Microducks remain an autonomous watch mode.
The ten Microducks walk and recover autonomously. A compact mobile panel keeps Pause, Drop again, Follow a duck and Back to Asimov within reach; detailed physics stats and source credits expand only when you want them.
Duck mode hides the Asimov driving controls. The ducks still use Pollen’s walking and stand-up policies in shared MuJoCo physics, including collisions. Recovery is attempted rather than replaced with an animation.
Asimov now starts ready to drive with the latest learned walking policy. Arrow keys, WASD and onscreen controls share the same simulation. If he falls, the recovery controller attempts to stand him up automatically and hands control back.
Scripted robot routines and the Microduck experiment remain separate. Walking and recovery benchmarks live under Diagnostics. Recovery is a physics-based attempt, not a reset or guaranteed success.
Menlo’s October 17 walking policy is now the default. OpenHorizon’s safe-fall and get-up policies add physical recovery tests, with clear simulation labels and source credits. No checkpoint selection needed.
Drive & test brings first-person controls and bounded tasks into one window. Ten Microducks share contacts with Asimov in a separate, clearly labelled legacy-policy experiment.
Touch feedback targeting, tablet layout, the downloadable SDK package and runtime dependencies have been updated. The browser demo works on its own; hosted Python SDK and LiveKit sessions await cloud configuration.
These are simulated results. Stair climbing is not validated, and object pickups remain scripted. Sources and credits