# When Less Is More: What 36 Lidar Experiments Taught Me About RL Agent Perception

![Lidar Configuration Example](https://res.cloudinary.com/dgz0aulnb/image/upload/v1762110339/blog_lidar_fov_360.png)

During my recent work on reinforcement learning agents navigating unknown environments, I observed something puzzling. My agents would often get stuck in circular scanning patterns—spinning endlessly as if searching for something. At first, I attributed this to poor reward shaping or exploration issues. But after running a systematic parameter sweep across lidar configurations, the data revealed something far more interesting about how RL agents actually perceive and navigate their world.

## The Hypothesis: More Beams, Better Navigation

Like many researchers, I started with an intuitive assumption: more sensor data should lead to better performance. If an agent can see more detail about its environment through additional lidar beams, surely it would navigate more effectively? This seemed especially reasonable given that modern autonomous vehicles use 64+ beam lidars specifically for their high resolution.

I also suspected that the circular scanning behavior was a fundamental limitation—perhaps agents needed to build complete mental maps before taking decisive action. But watching my agents spin in place for hundreds of steps suggested something else was happening.

## The Experiment: A Grid Search Across Perception Space

To test these assumptions systematically, I designed a parameter sweep across two key lidar dimensions:

**Number of beams:** 32, 64, 128, 256
**Field of view:** 120°, 180°, 360°

This gave me 12 distinct sensor configurations. For each configuration, I trained agents for 5000 epochs across 3 different random seeds—36 total experiments. The environment was a navigation task where agents must reach a target while avoiding obstacles, using only lidar distance measurements as input.

*360-degree lidar configuration showing full environmental awareness*

The key metric I tracked was termination progress completion probability—essentially, how often agents successfully reached their target without colliding.

## The Data: Counterintuitive Results

The results challenged both of my initial hypotheses:

![Lidar Parameter Sweep Results](https://res.cloudinary.com/dgz0aulnb/image/upload/v1762110340/blog_lidar_parameters_sweep.png)
*Left: Beam count has minimal impact on performance. Right: Field of view dramatically affects success rates.*

**Finding 1: Beam count doesn't matter**
Across all field-of-view configurations, agents with 32 beams performed nearly identically to those with 256 beams. The parallel coordinate plot on the left shows this clearly—performance lines are essentially flat across the beam count axis. More resolution provided no measurable benefit.

**Finding 2: Field of view is everything** 
The difference between 120° and 360° field of view was dramatic. Agents with full environmental awareness achieved ~80% success rates compared to ~70% for forward-facing sensors. This 10 percentage point improvement represents a significant performance gain.

![Training Curves by FOV](https://res.cloudinary.com/dgz0aulnb/image/upload/v1762110341/blog_lidar_training_curves.png)
*Learning curves show 360° FOV agents reach higher performance faster and more consistently*

## The Behavioral Insight: Eliminating the Circle Dance

The most striking discovery came from observing agent behavior during navigation. With 360-degree perception, agents virtually eliminated the circular scanning patterns that plagued narrow field-of-view configurations.

I recorded representative navigation episodes and the difference was immediately apparent. 360-degree agents moved with purpose—they might oscillate briefly when searching for optimal paths, but these oscillations led to decisive movement toward targets.

%[https://youtu.be/7ODVGc9gMQA]

This behavioral change suggests a fundamental shift in how agents process environmental information. With limited perception, agents must actively gather information through movement before acting. With full perception, they can make informed decisions immediately.

## Making Sense of the Data: The Key-Lock Mental Model

These results led me to reconsider how RL agents actually use sensory information. The beam count independence suggests agents aren't building detailed environmental maps. Instead, they appear to use lidar readings as pattern triggers—like keys that unlock specific behaviors when certain distance configurations are detected.

When an agent encounters a familiar pattern of beam distances, it activates a learned navigation behavior. The circular scanning emerges because agents with limited perception must "jiggle the key"—rotate until they find a distance configuration that matches a known safe path.

360-degree perception eliminates this search process. Agents can immediately access the full range of learned distance patterns without repositioning, leading to more direct and efficient navigation.

## Practical Implications

This data suggests several actionable insights for RL practitioners:

**1. Prioritize sensor coverage over resolution**
When designing perception systems for RL agents, field of view appears far more critical than sensor density. A sparse sensor array with full coverage outperforms dense forward-facing sensors.

**2. Question sensor-rich assumptions**
The robotics industry's preference for high-resolution sensors may not translate directly to RL agents. Our data shows diminishing returns from additional beams, suggesting computational resources might be better allocated elsewhere.

**3. Consider behavioral emergence**
The circular scanning wasn't a bug—it was an emergent search strategy for agents with limited perception. Understanding these behaviors as adaptive responses can guide both reward design and sensor configuration.

## What the Data Tells Us

After 36 experiments and thousands of training hours, the conclusions feel robust: for RL navigation agents in our environment, comprehensive environmental awareness matters far more than sensor resolution. The 10-point performance improvement from 360° perception, combined with the elimination of inefficient scanning behaviors, provides compelling evidence for this conclusion.

These findings challenge common assumptions about sensor design in RL and highlight how agents develop unexpected strategies when working with limited information. Sometimes in machine learning, as in many domains, less really can be more—as long as it's distributed in the right way.

---

*If you are interested in high-performance RL simulations, check out our [Reality project](https://github.com/DuaneNielsen/reality) - a simulation framework built on the Madrona engine.*