Researchers created an AI system to feed rainbow trout automatically
A new framework improves feeding decisions for rainbow trout by linking fish behavior data to LLM reasoning.
Researchers have developed a framework called THPL to help manage feeding for rainbow trout in recirculating aquaculture systems. The system uses cameras to track fish activity and converts these movements into data that a Large Language Model can understand.
How it works
Continuous spatiotemporal tokens provide necessary physical grounding for LLMs.
First, a tool called Fishsort tracks the fish to determine an Activity Coefficient. This score tells the system how energetic or hungry the fish appear to be based on their swimming patterns.
Next, a Hierarchical Behavior Encoder creates a digital representation of these physical movements. This data is combined with environmental information and expert feeding rules.
The system uses LoRA to fine-tune an LLM and applies a technique called counterfactual multimodal Direct Preference Optimization to ensure the AI makes safe, consistent decisions.
Performance results
The study found that this approach significantly improved decision-making accuracy. While a text-only AI model reached 33.33% accuracy, the full THPL framework achieved 96.67% accuracy in controlled tests.
Have you ever wondered if machines could understand animal behavior better than human observers?
Limits
This study specifically tested the system with rainbow trout in controlled recirculating environments. The researchers have not yet demonstrated how the framework performs with other fish species or in open-water settings.
Source: arXiv cs.AI