CLAE: New method steers multirobot behavior without retraining
Frozen policies gain new skills at inference time with closed-loop activation editing.
Real-world robots often need to adapt beyond their original training. Retraining risks catastrophic forgetting. To solve this, researchers from USC introduce CLAE (Closed-Loop Affine Activation Editing), a method that modifies a frozen policy's internal activations at inference time without altering base weights. CLAE trains a sparse autoencoder over policy activations, identifies behavior-relevant latent features via post-hoc probing, and learns a lightweight RL-based steering policy that applies affine edits to those latents online, adapting to robot state, environment, and multirobot context.
In simulations and physical tests with a multi-quadrotor navigation policy, CLAE demonstrated three capabilities: steering individual robot velocity profiles, preserving desired formations during navigation, and even generating entirely new behaviors like reducing exposure to surveillance cameras. This approach opens the door to flexible, safe robot swarms that can be repurposed on the fly without expensive retraining, potentially accelerating deployment in dynamic environments like warehouses or search-and-rescue.
- CLAE edits intermediate activations of a frozen policy using a sparse autoencoder and lightweight RL, avoiding catastrophic forgetting.
- Tested on multi-quadrotor navigation, CLAE enabled velocity control, formation preservation, and novel surveillance avoidance behaviors.
- Inference-time adaptation means robots can gain new skills without model retraining or additional data collection.
Why It Matters
Enables flexible robot behavior adaptation on the fly, reducing retraining costs and expanding deployment scenarios.