PaperlayerView exact revision on arXiv

Source-linked AI summary

From Visual Search to Movement Control: A Priority Field for Artificial Agents

Han Zhang, Zhong Cao

arXiv:2609.30704v1cs.RO

TL;DR

The paper asks whether priority-based computation used to explain spatial attention can also guide movement control. It models human visual search and extends the approach to a movement-direction priority field; agents trained more efficiently, generalized to unseen challenging scenarios, and showed history-driven goal anticipation.

  • Problem

    The paper asks whether priority-based computation used to explain human spatial attention can also account for movement control.

    Source: §1 · pp. 1–2

  • Method

    The authors model visual search with an integrated priority map and extend the computation to a movement-direction priority field for an agent.

    Source: §1 · p. 3

  • Results

    Priority-field agents trained more efficiently, handled unseen challenging scenarios, and produced human-like history-driven effects in reach–avoid tasks.

    Source: §1 · p. 4, §3 · p. 9

  • Takeaways & Limitations

    The results suggest that deciding where to look and where to move share an underlying structure that may support movement control in artificial agents.

    Source: §3 · p. 9

    The reach–avoid demonstration was limited to a two-dimensional simulation with range-only sensing and simple obstacles, requiring further development for richer perception and physical transfer.

    Source: §3 · p. 10

Abstract

from arXiv · show

Human spatial attention is widely conceptualized as being guided by a priority map that integrates perceptual salience, current goals, and past experiences. Here, we extend priority-based computation to movement control in artificial agents. We first introduce a lightweight model of visual search based on an integrated priority map. Trained on human saccades, it reproduced key behavioral patterns, including oculomotor suppression and history-driven selection. Extending the search model, we equipped an artificial agent with a priority field and evaluated its performance in a reach-avoid task that required reaching a goal destination while avoiding moving obstacles. Compared with alternative architectures, priority-field agents trained more efficiently and performed better in unseen, complex scenarios, even from simple demonstrations. Adding a simple memory mechanism also produced human-like, history-driven effects in anticipating the likely location of the upcoming goal. These findings suggest that priority-based computation may provide a promising foundation for movement control in artificial agents.

1. Introduction

The paper extends priority-based computation—which integrates salience, goals, and selection history—from visual search to movement control. It models human eye movements and evaluates an agent navigating toward goals while avoiding obstacles.

1. Introduction

  • The model treats visual search and spatial navigation as sharing principles: selection balances salient information, current goals, and past experience.In navigation, obstacle locations and approach speed can inform salience, while previous destinations can inform history.Source: §1 · pp. 1–3
  • The movement model represents salience, goal alignment, and recent destinations for each candidate direction, then uses a planner to convert the priority field into movement.Salience reflects collision threat, goal reflects destination alignment, and history reflects recent destinations.Source: §1 · p. 3
  • The visual-search model reproduced oculomotor suppression and history-driven effects when trained on human eye movements.Observers suppress salient singleton distractors and show first-fixation bias toward the previous target location.Source: §1 · pp. 2–3
  • Priority-field agents trained more efficiently and generalized better to unseen, complex reach–avoid scenarios while also anticipating likely goal locations from history.The agents were compared with an MLP and a transformer, and could learn from simple demonstrations.Source: §1 · p. 4

2. Results

The priority-field agent learned the demonstrated reach–avoid task efficiently and generalized to harder scenarios; adding memory produced anticipatory movement that tracked changing goal-location probabilities. A fitted priority-map model also reproduced key patterns of human visual search.

2.1. A priority map for visual search

  • On held-out participants, the search model’s mean negative log-likelihood was 1.40, outperforming the human-consistency baseline of 1.54.The model was fit to first fixations from 333 participants across 11 experiments, with 66 participants held out.Source: §2.1 · p. 4
  • For singleton distractors, the model predicted an 8.2% oculomotor suppression effect, close to the held-out participants’ 7.1% effect.Figure 2A presents the human and model effects.Source: §2.1 · p. 4, Figure 2 · p. 5
  • The model’s predicted suppression fell from 8.2% to 1.9% for low-salience distractors, and its 38.7% target-location priming effect was similar to humans’ 34.7%.Figure 2B compares salience conditions and Figure 2C compares human and model priming.Source: §2.1 · p. 5, Figure 2 · p. 5

2.2. A priority field for agent movements

  • The agent’s priority field combines directional collision threat, destination alignment, and recent destinations, which a planner network uses to produce movement.Training used behavioral cloning from a scripted expert over 20 episodes; the expert was memoryless, so the history weight was fixed at zero during training.Source: §2.2 · p. 5, Figure 3 · p. 6
  • After 500 epochs, the priority-field agent’s mean squared error was 0.004, compared with 0.111 for the MLP and 3.593 for the transformer.On the demonstrated task, it reached 60.60 goals/minute with 0.20 collisions/minute; the transformer reached 53.40 goals/minute with 0.23 collisions/minute.Source: §2.2 · pp. 6–8, Figure 4 · p. 7
  • At 50 obstacles, the priority-field agent reached 15.20 goals/minute versus 12.18 for the MLP and 9.70 for the transformer, while recording fewer collisions than both.Obstacle speed was held at the training level; Figure 4 compares goals and collisions across densities.Source: Figure 4 · p. 7, §2.2 · p. 8
  • At 4× obstacle speed, the priority-field agent reached 66.25 goals/minute, more than the MLP’s 51.65 and statistically indistinguishable from the transformer’s 64.95.Its collision rate was also lower than the MLP’s and statistically indistinguishable from the transformer’s.Source: Figure 4 · p. 7, §2.2 · p. 8
  • With memory enabled, the agent shifted −39.15 pixels toward the biased left region, then returned to an offset indistinguishable from the memoryless control when goals became unbiased.The biased block placed 70% of goals left for 90 trials, followed by 90 unbiased trials.Source: §2.2 · pp. 8–9, Figure 5 · p. 9

3. Discussion

The discussion reports that priority-field agents outperform architectures without explicit priority computation, and links their structured representation of salience, goal, and history to learning from little data and generalizing to unseen cases. It connects movement control with experience-driven cognition while qualifying the conclusions by the task and memory limits.

3. Discussion

  • The same priority-based structure links deciding where to look and deciding where to move.The authors propose that a model fit directly to behavior could quantify how salience, goal, and history shape movement in humans and other animals.Source: §3 · pp. 9–10
  • Priority-field agents outperformed models without explicit priority computation and generalized better to unseen cases while recovering behavior from little data.The authors relate this advantage to committing in advance to salience, goal, and history as the represented sources.Source: §3 · pp. 9–10
  • A leaky-accumulator history field connects movement effects to experience-driven cognition, including statistical learning, contextual cueing, and reward.The history field tracks previous goal locations; the authors note that the approach could also be used to study biological movement.Source: §3 · p. 10
  • The reach–avoid demonstration is limited to a 2D simulation with range-only sensing and simple obstacles, while its memory tracks only goal locations.The authors identify richer perception, physical-system transfer, and reward-based learning as areas for further development.Source: §3 · p. 10

4. Methods

The methods describe the human visual-search dataset and a priority-map model combining salience, goal match, and selection history. They also specify how the model was fit and evaluated against held-out participants and a human-consistency baseline.

4.1. Visual search task and dataset

  • The visual-search dataset contains 114,232 first fixations from 333 participants across 11 experiments, split 80/20 into training and held-out test sets.Canonical displays were reconstructed from trial parameters because original images were unavailable; trials with motion, onset, or low-salience distractors were excluded.Source: §4.1 · p. 10

4.2. The priority-map model

  • For each candidate item, the priority map is a weighted sum of salience, goal match, and selection history.The maps measure perceptual salience, similarity to target-defining features, and influence from previous target locations.Source: §4.2 · p. 11
  • Salience is the item’s color distance from the display average, while the goal map equally weights color and shape match to the search template.The salience score is scaled to [0, 1]; the goal map ranges from −1 to 1, allowing target-feature upweighting and distractor-feature downweighting.Source: §4.2 · p. 11
  • Selection history uses a leaky accumulator that decays traces at item locations and updates the trace at the current target location.The update rate ηH controls updating, while 1 −ηH controls decay; higher ηH biases the trace toward recent events.Source: §4.2 · p. 11
  • A softmax over item priorities gives first-fixation probabilities, and one item is sampled from that distribution to predict each trial’s fixation.Higher-priority items receive higher landing probabilities.Source: §4.2 · p. 11
  • The four free parameters were estimated by maximum likelihood with Adam for 300 full-batch iterations at a learning rate of 0.05.The parameters were the salience, goal, and history weights and the memory-update rate; wG > 0 and 0 < ηH < 1 were constrained.Source: §4.2 · p. 11
  • Evaluation compared held-out participants’ first-fixation negative log-likelihoods with a human-consistency baseline estimated from training-set fixation probabilities.Model and baseline NLLs were averaged within participant and compared using paired t-tests, with Cohen’s d as effect size.Source: §4.2 · p. 11

4.3. Behavioral-signature analyses

The analyses test whether the fitted model reproduces three behavioral signatures in held-out trials: oculomotor suppression, salience dependence, and target-location priming.

4.3. Behavioral-signature analyses

  • The analyses measure singleton suppression, salience dependence under a low-salience counterfactual, and target-location priming by comparing repeated versus changed target locations.Participant-level landing probabilities were compared with paired t-tests using α = 0.05 and Cohen’s d as effect size.Source: §4.3 · p. 12

4.4. The priority-field model

The model assigns priority across 360 movement directions by combining salience, goal, and history fields. A planner maps the resulting field to movement.

4.4. The priority-field model

  • The model sums salience, goal, and history contributions across 360 candidate directions to form the priority field.A planner network maps this field to the movement vector.Source: §4.4 · pp. 12–13
  • Obstacle proximity and looming signals are transformed into a salience field for each movement direction.A one-hidden-layer network combines the two channels; the distance-to-closeness transformation has learned, direction-specific sensitivity.Source: §4.4 · p. 12
  • The goal field scores each direction by its angular alignment with the goal, then modulates that score using a learned function of goal distance.Alignment is the cosine of the angle between the candidate direction and the direction to the goal.Source: §4.4 · p. 13
  • Goal history is accumulated on an 8 × 8 grid, with past weights decaying and the current goal cell incremented after each goal onset.The grid is transformed into egocentric coordinates by treating each cell as a weak goal; the resulting history field is active only when the current goal is not visible.Source: §4.4 · p. 13
  • A planner multilayer perceptron converts the 360-dimensional priority field into a two-dimensional movement vector.The planner has two hidden layers of 360 rectified linear units.Source: §4.4 · pp. 13–14

4.5. Reach–avoid task

The reach–avoid task tests continuous movement toward a goal while navigating moving obstacles in a bounded arena.

4.5. Reach–avoid task

  • The agent continuously navigates an 800×800-pixel arena with 10 moving obstacles and one goal destination.Obstacles are same-sized disks that move diagonally at 1 to 3 pixels per step along the axes, bounce off walls, and pass through one another.Source: §4.5 · p. 14
  • The agent senses the nearest obstacle or wall distance in each direction for the current and previous frames, plus a vector to the goal.Its maximum speed is 10 pixels per step, and a goal counts as reached within 30 pixels.Source: §4.5 · p. 14
  • A new goal appears after the agent reaches one or fails to reach it within 6 seconds.The task therefore continues across successive goal destinations.Source: §4.5 · p. 14

4.6. Training

The agent learns movement by behavioral cloning from a scripted expert, with history disabled during training because the demonstrations contain no history-driven behavior.

4.6. Training

  • The expert generated 12,000 state-action pairs by completing 20 independent episodes of the continuous task.The episodes used 10 obstacles and unbiased goal locations.Source: §4.6 · p. 14
  • Behavioral cloning minimizes mean squared error between the agent’s predicted movements and the expert’s movements.Training used Adam for 500 epochs with a batch size of 64 and a plateau-based learning-rate schedule.Source: §4.6 · p. 14
  • The salience, goal-gain, distance-sensitivity, and movement-planner components were jointly optimized, while history weight was fixed at zero.The expert was memoryless, so its demonstrations contained no history-driven behavior; history weight and update rate were set separately afterward.Source: §4.6 · p. 14

4.7. Alternative models

The priority-based model was compared with MLP and transformer alternatives that received the same input but lacked explicit priority-based computation.

4.7. Alternative models

  • The MLP baseline maps the 722-length input through hidden layers of 360 and 180 rectified linear units to movement.Its architecture is 722 →360 → 180 →2.Source: §4.7 · p. 14
  • The transformer embeds input values as 16-dimensional tokens, then applies a two-layer encoder with four attention heads and feedforward width 64.Token representations are averaged across the embedding dimension before a linear layer maps the resulting 722-length vector to movement.Source: §4.7 · p. 14
  • Both alternative models used the same demonstrations and training regime as the priority-based model.They received identical input but had no explicit priority-based architecture.Source: §4.7 · p. 14

4.8. Evaluation

The evaluation compares trained agents on the training configuration, more challenging obstacle scenes, and faster-obstacle scenarios. Performance is measured by goal-reaching speed and collision rate across repeated random seeds.

4.8. Evaluation

  • Agents were tested on the training configuration, on scenes with 20–50 obstacles, and at obstacle speeds from 1.5× to 4× training speed.The latter two evaluations varied obstacle count or speed while holding the other factor at its training level; each scenario ran for 6,000 steps.Source: §4.8 · p. 15
  • Goals reached per minute and collisions per minute measured speed and safety, respectively, across 20 random seeds per model and scenario.Reported means used bootstrapped 95% confidence intervals.Source: §4.8 · p. 15
  • Paired t-tests used α = 0.05, with Cohen’s d reported as the effect size.Source: §4.8 · p. 15

4.9. Statistical learning

The history parameters were recovered to two decimal places from goal-free movement behavior. The experiment tested whether grid memory shaped anticipatory movement toward a spatially biased goal location.

4.9. Statistical learning

  • Trials began with the agent at arena center among 10 obstacles and no visible goal; after 50 steps, a goal appeared in one of two symmetric wedges.In biased trials the goal appeared on the left with 70% probability; unbiased trials assigned equal probability to either side.Source: §4.9 · p. 15
  • Grid-memory and memory-disabled agents were compared on anticipatory horizontal offset across 90 biased trials followed by 90 unbiased trials.The procedure was repeated for 20 random-seed runs; negative offsets indicate movement toward the high-probability region.Source: §4.9 · p. 15
  • The fitted history parameters recovered their generating values to two decimal places when network weights were held fixed.Only wH and ηH were estimated by minimizing mean squared error between predicted and recorded goal-free movements.Source: §4.9 · p. 15
Loading 2609.30704v1…