Preprint · 2026

Behavior Uncloning

Distilling Mode Redirection into Policy Weights without Inference-Time Steering

Hao Wang^1,†, Jiuzhou Lei¹, Dayou Li¹, Bangya Liu², Minghui Zheng¹, Manling Li³, Ruohan Zhang^3,4, Zhiwen Fan¹

¹Texas A&M University ²University of Wisconsin ³Northwestern University ⁴Stanford University † Project Lead

Paper Code Watch the video Interactive demo

Real robot Instruction“pass the knife”

Original policytwo modes

✗ blade toward the arm

✓ handle first

MoRE-edited policyone mode

✓ handle first

Behavior cloning copies every mode in the demonstrations, including the unsafe one. MoRE edits it out of the weights offline; the edited policy keeps the original inference path.

Video

Behavior Uncloning in 80 seconds.

Abstract

Distilling mode redirection into policy weights.

Behavior-cloned policies often learn multiple behavior modes from demonstration datasets, including modes that are unsafe, inconvenient, or otherwise undesired at deployment. Standard remedies such as data curation and inference-time steering either require full retraining or add extra computation to every control step.

MoRE redirects policy rollouts toward desired behavior modes through a short uncloning step. A behavior-mode classifier supplies a differentiable redirection signal during editing, while a retain loss preserves desired-mode competence. After editing, the classifier is no longer needed: the updated policy runs through the same inference path as the original policy.

+44pp

average deployment success over the original mixed policy.

simulated and real-world tasks across manipulation and navigation.

<500

editing gradient steps per target setting in our evaluations.

extra inference-time steering modules after the edit.

Method

Differentiable mode-redirection editor.

MoRE treats behavior mode control as a policy editing problem. It uses mode-labeled examples to train a mode classifier, then distills the desired redirection into the policy weights so the final policy can be deployed without an auxiliary steering module.

The policy’s distribution over behavior modes during the edit. A temporary classifier supplies the redirection signal, a retain loss keeps the desired mode intact, and the classifier is dropped before deployment.

Train a mode classifier

Fit a K-way classifier on features cached from the original mixed policy.

→

Redirect undesired samples

Use differentiable policy features to move probability mass toward the desired mode set.

→

Preserve task competence

Keep the original imitation loss on desired-mode samples and deploy the edited policy by itself.

L_MoRE = L_retain(desired samples) + γ · L_redirect(gated undesired samples)

Real-Robot Deployment Results

Mode edits transfer to physical robot tasks.

Across four real-robot tasks with a Franka Research 3 arm, MoRE shifts rollouts toward requested behavior modes while preserving task completion.

Knife handover

SR: 45.0 → 70.0 (+25.0). Clip: original vs. MoRE-edited rollouts handing the knife to a person.

Task	Mode change	Original Avg SR	MoRE Avg SR	SR Gain (pp)	MoRE TCR
Place-Knife	2 → 1	40.0%	80.0%	+40.0	80.0%
Place-Bottle	3 → 1	31.7%	96.7%	+65.0	96.7%
Block-Obstacle	2 → 1	47.5%	92.5%	+45.0	92.5%
Knife-Handover (VLA)	2 → 1	45.0%	70.0%	+25.0	85.0%

Values report average real-robot SR over target modes. Place-Bottle also supports a 3 → 2 edit, improving SR from 50.0% to 100.0%.

Simulation

Choose a mode. The policy follows.

Each simulated task admits several behavior modes. The original policy spreads its rollouts across all of them; a MoRE edit keeps the target mode you pick and removes the rest, with the same inference path.

Click a mode to edit toward it

MoRE vs. adaptation and steering baselines

Deployment success rate (SR, %) averaged over target modes, Diffusion Policy benchmark. DynaGuide steers at inference time and runs at 8.47× the original latency; CPL and MoRE edit the weights and keep the original latency. The dashed marker is the filtered-retrain reference, a policy retrained from scratch on mode-filtered demonstrations.

Task	Modes	Original Avg SR	MoRE Avg SR	SR Gain (pp)	MoRE TCR
Push-T	left or right	50.0%	98.3%	+48.3	98.7%
Push-Wall	left or right	37.0%	80.7%	+43.7	80.7%
Push-Pillars	four routes	21.5%	72.3%	+50.8	80.7%
Quadruped	left or right	50.0%	84.3%	+34.3	84.7%

Values report average deployment success rate (SR) and task completion rate (TCR) over target modes.

Supplementary video

Rollout grids across the evaluation suite.

The supplementary video walks through the evaluation suite: rollout grids for each simulated and real-robot task under the original and MoRE-edited policies, and the compared baselines.

Team

Authors.

Citation

Cite this work.

If you find Behavior Uncloning useful, please consider citing:

@misc{wang2026behavioruncloning,
      title={Behavior Uncloning: Distilling Mode Redirection into Policy Weights without Inference-Time Steering},
      author={Hao Wang and Jiuzhou Lei and Dayou Li and Bangya Liu and Minghui Zheng and Manling Li and Ruohan Zhang and Zhiwen Fan},
      year={2026},
      eprint={2606.29201},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2606.29201},
}