Proxy Policy Steering

Chuanruo Ning*, Tianrui Wang*, Wei-Chiu Ma, Kuan Fang * Equal contribution     Equal advising
Cornell University Cornell University
CoRL 2026
We propose Proxy Policy Steering (PPS), which adapts a frozen base policy by training two lightweight proxy policies. By adding their velocity-space residual to the base in the flow-matching process, PPS steers the action distribution toward task behaviors while preserving the base policy's general priors.
Abstract

Generalist robot policies carry broad manipulation priors from large-scale data, but specializing them to a new task remains the deployment bottleneck. This requires eliciting task-specific behavior from limited demonstrations without degrading their broad capabilities. We introduce Proxy Policy Steering (PPS), an inference-time adaptation method that resolves this challenge by training two lightweight proxy policies whose calibrated velocity-space difference steers the frozen base sampler. A reference proxy models the frozen base's behavior on target-task observations, and a task proxy, initialized from the reference, captures how this behavior changes under task supervision. Their difference forms a calibrated velocity-space residual that steers the frozen base sampler at every denoising step. We identify the conditions under which this residual isolates the change induced by task supervision, and validate them empirically. Because the base is never directly modified, its broad capabilities remain available at inference, including behaviors such as recovery from failure that the demonstrations themselves do not exercise. Adaptation requires only forward velocity predictions from the base, making PPS lightweight to train and applicable even without access to the base's parameters. On 8 real-world and 4 simulation manipulation tasks, PPS lifts the state-of-the-art π0.5 base policy by 55% absolute success rate on average, with zero-to-one gains on tasks the base never solves, while preserving the base's broad capabilities. PPS outperforms LoRA fine-tuning, from-scratch specialists, residual policies, and prior inference-time steering methods.

Method

Our key insight is that effective adaptation should neither steer only within the base's existing distribution nor replace the base with a narrowly trained specialist. Instead, PPS estimates the direction in which task supervision would shift the base's action distribution, leaving its weights and broad priors untouched.

To estimate that direction without access to base gradients, PPS trains two lightweight proxy policies and takes their difference in velocity space.

a log πbase
γa log πref
+
γa log πtask
a log πPPS
base distribution
πbase
Distill
ref distribution
πref
Fine-tune
task distribution
πtask
pps distribution
πPPS
rollout 1
Steer to task behavior
rollout 2
Recover via base prior
rollout 3
Agree with base

At inference, PPS integrates the velocity predictions vi of the three policies πi in each flow-matching step, with coefficient γ representing the steering strength:

vPPS = vbase + γ · (vtask vref)

Here we show three representative cases: a state where the base acts differently from the demonstrations (residual steers toward the task behavior), a state not covered by the demonstrations (residual cancels, base prior recovers), and a state where the base already agrees with the demonstrations (residual vanishes).

Effective Steering

We evaluate PPS on 8 real-world and 4 simulation manipulation tasks, spanning long-horizon goals, articulated and deformable objects, and diverse skills. PPS delivers consistent gains over the base model and outperforms fine-tuned models, from-scratch specialists, and prior steering baselines, with zero-to-one gains on tasks the base never solves on its own.

Three bar charts — Real-world, Simulation, and π0 transfer — comparing average success rates of π0.5 base, Specialist, Residual, DSRL, LoRA, and PPS. PPS leads every panel.
PPS lifts π0.5 from 22% to 79% in the real world and from 12% to 64% in simulation, and lifts π0 from 5% to 55% with the same recipe.

Select a task below to compare PPS rollouts with the frozen base policy and the strongest baselines on the same task.

Coffee Brewing
Base: 0/10 PPS: 5/10
PPS
Success
Base (π₀.₅)
Can't open the lid
LoRA
Overfits after a near miss
PPS
Success
Base (π₀.₅)
Places instead of inserting
LoRA
Continues a wrong insertion
PPS
Success
Base (π₀.₅)
Can't fold
LoRA
Overfits, tosses the jeans
PPS
Success
Specialist
Overfits to the demonstrations
LoRA
Tips the box, can't recover
PPS
Success
Base (π₀.₅)
Releases before wiping
LoRA
Drags before the tissue is out
PPS
Success
Base (π₀.₅)
Can't open the drawer
LoRA
Stows before opening
PPS
Success
Specialist
Overfits to the pressing motion
LoRA
Overfits to the pressing motion
PPS
Success
Base (π₀.₅)
Imprecise insertion
LoRA
Overfits to the insertion motion
PPS
Success
Base (π₀.₅)
Misses the placing
LoRA
Misses the grasp, can't recover
PPS
Success
Base (π₀.₅)
Picks egg before opening lid
LoRA
Overfits to incorrect placing mode
PPS
Success
Base (π₀.₅)
Can't pick up the teapot
LoRA
Misses the grasp, can't recover
PPS
Success
Base (π₀.₅)
Can't open the lid
LoRA
Misses the grasp, can't recover
Robust Execution

Because PPS leaves the base model untouched, the combined policy inherits its broad priors and recovers from failures it never saw in the demonstrations. LoRA fine-tuning, by contrast, overfits to the handful of demonstrations and overwrites parts of these priors, breaking down under the same perturbations.

LoRA
PPS
LoRA
PPS
Effect of Steering Strength

Drag the steering strength γ and watch the trade-off between base priors and task guidance.

Cross-Modality Steering

Because the proxies are separate networks, they can take inputs the base never receives. PPS routes extra modalities such as point cloud or audio into the frozen, RGB-only base through the task proxy: the point cloud locates a textureless object, and audio tells the ringing phone from the silent one. This makes PPS a lightweight way to add new sensors without retraining the base.

Additional Modality
PPS Policy Rollout
Success Rate
point cloud
Marble task success rate: PPS reaches about 82%, ahead of Specialist, LoRA, LoRA-MM, DSRL, and the base policy.
audio
Phone task success rate: PPS reaches about 86%, ahead of LoRA-MM, Specialist, LoRA, DSRL, and the base policy.
Legend: base policy, Specialist, DSRL, LoRA, LoRA-MM, PPS.
Analysis of PPS

Through a series of analyses, we study what makes PPS effective and where it still falls short.

Success rate versus steering scale: PPS peaks near 0.4 to 0.6, above the specialist and base baselines.

Steering strength. The coefficient γ trades base priors against task steering. Success peaks for γ in (0.4, 0.6), well above both baselines.

Data scaling on Tea and Capsule: success rate versus number of demonstrations from 50 to 200 for Specialist, LoRA, and PPS.

Data scaling. On the two lowest-success tasks (Tea and Capsule), PPS leads at every demonstration budget from 50 to 200, with its largest margin in the low-data regime. The from-scratch specialist improves steadily and catches up near 200 demonstrations.

Failure-mode breakdown for Specialist, base, and PPS by root cause; PPS has far fewer total failures.

Failure modes. PPS fails far less often than the specialist and base, cutting incorrect-mode errors 93% versus the base and out-of-distribution errors 87% versus the specialist.

Proxy-training ablation: w/o ref, w/o vel, and w/o tune all trail the full method.

Proxy training. On-policy distillation, velocity-level supervision, and reference initialization each ensure the two proxies differ by task supervision alone, and their difference is an informative task signal.

BibTeX
@inproceedings{ning2026proxy,
  title     = {Proxy Policy Steering},
  author    = {Ning, Chuanruo and Wang, Tianrui and Ma, Wei-Chiu and Fang, Kuan},
  booktitle = {Conference on Robot Learning (CoRL)},
  year      = {2026}
}