CoRL 2026

Artificial Foveated Perception for Mitigating Shortcut Learning in Robotic Foundation Models

Xiatao Sun1†, Yuan Zhuang2, Mateo Sanchez Lopez Negrete1, Matei-Victor Coldea1, Chen Liang1, Haoyang Zhang3,5, Che Liu4,5, Ziyao Zeng1, Shawn Li5, Qian Wang1, Fei Miao2, Daniel Rakita1
1Yale University  ·  2University of Connecticut  ·  3Peking University  ·  4Imperial College London  ·  5Digients
Corresponding author: xiatao.sun@yale.edu

Overview

Fine-tuning teaches a policy what to do. It does not say where to look.

Camera
AFP mask
Masked view

AFP on a manipulation sequence. Values near 1 mark the objects and robot parts the task depends on.

Robotic foundation models still need task-specific fine-tuning before they work reliably in a new setting, and that fine-tuning supervises actions only. Nothing constrains which visual evidence the policy adapts to, so it can settle on cues that merely predict the demonstrations: background texture, lighting, object co-occurrence, camera artifacts. These shortcuts hold in the training scene and fail under distribution shift.

AFP is a small, policy-agnostic module that takes the same RGB frames and instruction as the policy and predicts a continuous mask over the task-relevant objects, the robot and other action-critical regions. During fine-tuning the mask supervises the policy's attention over image tokens. The architecture is unchanged, and once fine-tuning ends the policy runs on the raw observation stream without AFP. Across SmolVLA, OpenVLA, π0.5 and Motus on eight simulated tasks, and π0.5 on five real-robot tasks, this raises success under visual distractors, matches or improves in-distribution success on every task, and shortens fine-tuning on long-horizon tasks.

Method

A mask predictor, and a loss that uses it

Given a sequence of frames I1:T and a task description 𝒯, AFP predicts one dense relevance mask Mt[0,1]H×W per frame.

Input
RGB frames
The policy's own camera stream, one short window at a time.
Input
Task description
The same language instruction the policy receives.
Artificial Foveated Perception
Two light pathways, conditioned on language
Feature path
A MobileNetV3-Large backbone gives a multi-scale feature pyramid. Each level is projected, flattened into visual tokens and passed through a deformable transformer encoder.
Query path
A few learned region queries are decoded against the visual memory with deformable cross-attention. They act as compact descriptors of the task-relevant objects and the end-effector. Temporal modules on both paths keep masks stable across frames.
Language
A frozen CLIP text embedding drives a FiLM adapter that rescales and shifts the decoded queries, with a scalar gate.
Mask head
Dense multi-scale features are fused into a high-resolution map. The primary query is matched against every location and a sigmoid gives the mask.
Output
Task-conditioned mask
Near 1 on relevant objects and robot parts, near 0 elsewhere.
Use
Attention supervision during fine-tuning
Pooled to the policy's image-token grid and compared with the policy's attention.

Language conditioning

The instruction reweights the region queries through a FiLM adapter instead of being concatenated as tokens, and the text encoder stays frozen.

Q~t = Qt + c𝒯 ( γ𝒯Qt +β𝒯 ) , (γ𝒯,β𝒯,c𝒯) = gψ(e𝒯)

Per-channel scale γ in [−1, 1], shift β, and gate c in [0, 1], all from the CLIP embedding e𝒯.

The grounding loss

During fine-tuning we read the policy's attention over image tokens, average it over trainable layers, heads and query positions, and compare the result with the pooled AFP mask. Both sides are normalized, so the loss compares distributions, not attention mass. The policy is trained on =act+λAFP.

AFP = 1B b=1B a¯bia¯b(i) wbiwb(i) 22

a¯ is the averaged policy attention over image tokens and w the AFP mask pooled to the same token grid.

Projected gradients

The grounding signal should steer action learning, not compete with it. Following PCGrad, when the AFP gradient points against the action gradient we drop its conflicting component and update with g=gact+λgAFPPC. Without the projection the total loss falls at first, then spikes after an onset point a little past the halfway mark of training. With it, training stays smooth.

11,152
projections triggered over 20k fine-tuning steps
gAFPPC = { gAFP gAFPgactgact22 gact if gAFPgact<0 gAFP otherwise

In action

Where the policy looks

Attention attribution during out-of-distribution rollouts, with distractors added to the scene. Left: direct fine-tuning. Right: fine-tuning with the AFP loss. Same model, same demonstrations, same training budget.

Simulation, MimicGen

Direct fine-tuning
With AFP

Real robot, π0.5 on an i2RT YAM Arm

The distractors are unused 3D-printed furniture parts, socks that never appeared in training, or an opened cabinet door. Clips end when the episode ends, so the two sides differ in length.

Direct fine-tuning
With AFP

AFP masks on real robot data

Top-camera sequences with the predicted mask applied. The mask follows the objects being manipulated and the gripper, and drops the rest.

Drawer Assembly
Lamp Assembly
Table Assembly
Cloth Folding

Results

Grounded attention, and policies that survive distractors

Attention against human task masks

We compare each policy's attention attribution with human-labeled task masks over 10 in-distribution and 10 out-of-distribution rollouts per task, averaged across the eight simulated tasks. Soft-IoU measures overlap. EMD measures how far the attention mass sits from the mask. Direct fine-tuning leaves attention on background and distractors for every backbone we tried.

0.93
Soft-IoU for π0.5 with AFP, in distribution
from 0.17 with direct fine-tuning
0.04
EMD for π0.5 with AFP, in distribution
from 4.39 with direct fine-tuning
Direct fine-tuningWith AFP
Soft-IoU
Higher is better
Earth mover's distance
Lower is better

Simulation success rates

Eight MimicGen tasks, 500 demonstrations and 20k fine-tuning steps each. The out-of-distribution setting keeps the objective and adds three to five unrelated RoboSuite objects with random materials to the tabletop.

Direct fine-tuningWith AFP

Real robot

π0.5 on an i2RT YAM Arm, five tasks, 300 demonstrations and 50k fine-tuning steps each, 30 rollouts per setting. Drawer, Lamp and Table Assembly come from FurnitureBench. AFP improves every task in both settings, with the largest gains out of distribution.

In distribution
Out of distribution

Drawer Assembly rollouts in the training scene and with spare 3D-printed parts from the same furniture family added as distractors.

Direct fine-tuningWith AFP

Faster on long-horizon tasks

AFP also shortens fine-tuning. On short tasks such as Square D1 and Coffee D1, both runs level off at about the same step and the gain is modest. On Kitchen D1, Coffee Preparation D1 and Three Piece Assembly D1, which have longer episodes and several interaction phases, the AFP-trained policy reaches high success well before the directly fine-tuned one. The gain seems to grow with the share of the observation stream that is incidental to the task.

Limitations

AFP depends on task-relevance annotation: humans label masks, AFP learns from them, and only then supervises the policy. That makes the supervision explicit and reusable, but adds labeling cost and possible bias. AFP can fail when relevance is ambiguous, heavily occluded, or outside the distribution of the mask dataset. Two directions follow: scale the dataset with the annotation tool, or learn relevance from interaction, temporal consistency, counterfactual scene changes and policy failures.

Average episode length
Steps per task, averaged over episodes

Labeling a real-robot episode, played at 5x. Point prompts on the first frame, a dozen instance layers, then propagation through the sequence.

Data and annotation

Soft masks, and the tool that made them

AFP is trained as a supervised mask predictor. Its targets are graded relevance fields, not binary segmentations: values fade around object boundaries, contact regions and partial occlusions. Standard annotation tools assign each pixel to one object, so we built our own.

Open an episode folder in the browser and place positive and negative point prompts on a key frame. Separate semantically different regions into instance layers.
SAM 2 turns the prompts into a key-frame mask for each instance.
The merged mask initializes MatAnyone, which propagates it through the whole sequence and writes a per-frame foreground image and alpha mask.

With this tool we labeled frames from Open-X Embodiment, DROID, MimicGen, LIBERO and our own scenes. Each training example is an RGB frame, an instruction and a continuous mask. The training dataset is on Hugging Face.

Citation

Cite this work

@article{sun2026artificial,
  title={Artificial Foveated Perception for Mitigating Shortcut Learning in Robotic Foundation Models},
  author={Sun, Xiatao and Zhuang, Yuan and Negrete, Mateo Sanchez Lopez and Coldea, Matei-Victor and Liang, Chen and Zhang, Haoyang and Liu, Che and Zeng, Ziyao and Li, Shawn and Wang, Qian and others},
  journal={arXiv preprint arXiv:2607.10655},
  year={2026}
}