Enlarge figure
Real-world robots need to adapt their behavior beyond the envelope of their pre-trained policy. Policy finetuning or retraining are options, but they risk catastrophic forgetting, degrading the pretrained policy’s base performance. To combat this, we introduce CLAE: Closed-Loop Affine Activation Editing, an inference-time framework for steering the behavior of a frozen policy by editing intermediate activations while keeping the base policy weights and downstream action head untouched. CLAE approaches behavior steering as a closed-loop problem whose outputs edit policy activations that adapt online to the robot state, environment, target behavior, and multi-robot context. It trains a sparse autoencoder over frozen-policy activations, selects behavior-relevant latent features via post-hoc probing, and learns a lightweight RL-based steering policy that applies state-dependent affine edits to selected latents during inference. We validate CLAE on a frozen multi-quadrotor navigation policy trained to perform a single task: navigating robots to a set of goal locations while avoiding obstacles. Through extensive simulations and physical tests, we show that while navigating to their goal positions, CLAE can 1. steer individual robot behavior by controlling each robot’s velocity profile; 2. coordinate multirobot behavior by preserving a desired formation, and 3. produce entirely new behavior wherein robots are required to reduce their exposure to surveillance cameras in the environment.
Follow a desired velocity profile while navigating to the goal.
Move as a team while negotiating obstacles.
Reduce cumulative exposure to surveillance cameras.
Watch four quadrotors maintain formation around obstacles.
02 / Method
CLAE treats behavior steering as a closed-loop problem. Edits adapt to the robot state, the environment, the target behavior, and neighboring robots at each policy step.
Follow an activation through the editing pipeline.
Step 1 of 6
The pretrained navigation policy maps observations to robot actions. Its weights, architecture, and downstream action head stay fixed throughout CLAE training and deployment.
Step 2 of 6
At each policy step, intercept the activation at a chosen internal layer. This creates an editing interface inside the policy, before the remaining frozen layers compute the robot’s action.
Step 3 of 6
A sparse autoencoder trained on frozen-policy activations maps the intercepted activation into a sparse latent representation. This provides a feature basis for selective intervention.
Step 4 of 6
Post-hoc probes use rollout-derived behavior metrics to select a compact latent set for each target behavior. Only these selected features can be edited; the remaining SAE latents pass through unchanged.
Step 5 of 6
A lightweight RL steering policy uses the current context to predict multiplicative gains and additive offsets for the selected features. These edit parameters change online with the robot state, environment, target behavior, and teammates.
Step 6 of 6
The SAE decoder maps the edited latents back to a policy activation. Reinsert it into the frozen network, let the original action head produce the robot action, and repeat the editing process with the next observation.
Scroll the diagram sideways or tap it to enlarge.
Intercept an intermediate activation of the frozen policy. A sparse autoencoder maps it into a sparse latent basis.
Post-hoc behavior probes select relevant latents. A lightweight, behavior-specific RL steering policy predicts their affine gains and offsets online.
Decode the edited latents and reinsert the activation. The original frozen downstream layers produce the robot’s action.
ℓ′j = mj ℓj + cj
Affine edits apply only to the selected latent features. Unselected latents pass through unchanged; the base policy weights and action head remain fixed.
Training and deployment. The SAE and behavior-specific steering policies are trained separately from the frozen base policy. At deployment, the steering policy chooses edits online; it does not update the base policy’s weights.
03 / Demonstrations
Compare the original navigation policy with CLAE in the same simulated environments. Each task uses its own steering policy and behavior objective.
Individual behavior
CLAE adapts each robot’s motion to a time-varying reference. Trajectory color indicates velocity-tracking error; lighter trajectories indicate closer tracking.
CLAE adapts each robot’s motion to a time-varying reference.
Both policies navigate the same scene. CLAE follows the requested velocity profile more closely.
Collective behavior
The frozen base policy navigates independently. CLAE uses teammate context to steer robots toward a target square formation. Colors identify robot teams.
The frozen base policy navigates independently. CLAE uses teammate context to steer robots toward a target square formation.
With square-structured starts and goals, CLAE keeps the team closer to the desired geometry around obstacles.
From unstructured starts, CLAE brings robots toward the target square as they navigate to their goals.
New behavior
Camera avoidance was absent from the base policy’s training objective. Blue regions indicate camera coverage, and trajectory color denotes speed.
Camera avoidance was absent from the base policy’s training objective.
When low-exposure paths exist, CLAE routes around camera regions or crosses them faster to reduce cumulative exposure.
With dense coverage, CLAE prioritizes lower exposure over goal progress. In this example, some robots stop short of their goals.
Enlarge figure
04 / Results
Reduction factors in simulation relative to the frozen base policy. See results
For each task, results are averaged over the same 1,000 distinct simulated environment configurations. Lower values indicate better behavior-specific performance.
| Metric | Base Policy | CLAE | Reduction factor |
|---|---|---|---|
| Velocity error (m/s) ↓ | 0.322 | 0.085 | ≈3.8× |
| Formation error ↓ | 0.280 | 0.086 | ≈3.2× |
| Camera exposure (watchability) ↓ | 3596.6 | 1151.3 | ≈3.1× |
Formation error measures deviation from the target square geometry. Watchability counts cumulative robot–camera visibility events per episode.
Enlarge figure
The paper includes additional activation-editing baselines, residual-action learning, diffusion-policy steering, and design ablations. Read the full evaluation
05 / Real world
We deploy CLAE zero-shot from simulation on four Crazyflie 2.1 quadrotors to steer the team toward a desired formation while navigating around obstacles.
The base policy reaches the goals by splitting around the middle obstacle. CLAE keeps the team closer together. The clips have different durations and play independently at their original speeds.
Entire pipeline onboard
Transfer from simulation
Coordinated formation control
The frozen policy, sparse autoencoder, and steering policy run on each robot’s STM32 microcontroller, alongside onboard state estimation and inter-robot communication.
06 / Citation
If CLAE is useful to your research, please consider citing the paper.
@misc{das2026steeringmultirobotbehavior,
title = {Steering Multirobot Behavior via Closed-Loop Affine Activation Editing},
author = {Satyajeet Das and Darren Chiu and Shashank Hegde and Gaurav S. Sukhatme},
year = {2026},
eprint = {2606.11489},
archivePrefix = {arXiv},
primaryClass = {cs.RO}
}