Custom Recurrent RL Policies

Academic

Structures found in this family (2)

Linear Direction, Limit cycle (stable periodic attractor)

By model (3)

Custom frozen multi-quadrotor navigation policy · research-scale

Linear Direction

GRU policy (PPO, Procgen Jumper)

Limit cycle (stable periodic attractor)

Mamba policy (PPO, Procgen Jumper)

Limit cycle (stable periodic attractor)

Observations (2)

SAE latents in a frozen multirobot policy are steerable via closed-loop affine edits

Trained recurrent (RNN/GRU/Mamba) policies' hidden states spontaneously converge to a stable, low-dimensional closed-orbit limit cycle that is contractive under finite-time Lyapunov analysis, and selectively randomizing only the top CCA-aligned dimensions between neural state and behavior causally collapses navigation performance back to baseline while randomizing all other dimensions preserves it

Papers

Steering Multirobot Behavior via Closed-Loop Affine Activation Editing (2026), Unraveling the Hidden Dynamical Structure in Recurrent Neural Policies (2026)