Implementation of causal experiments
This repository supports the study "Dribble in the Mind: Exploring Causality with Cognitive Soccer Agents", a project investigating causal reasoning in robotic soccer simulations using a hybrid cognitive architecture. The simulation features a goalkeeper agent attempting to infer the motion of a ball and the intentions of a striker using only partial sensory information.
The code for the work "Dribble in the Mind: Exploring Causality with Cognitive Soccer Agents" is present in branch nao.
"Dribble in the Mind" Banner: Available Here
Run first the scene present in scenes using CoppeliaSim_Edu_V4_3_0_Ubuntu20_04
Then run the java project CST_Causality_App
The project demonstrates a hybrid cognitive architecture that fuses neural perception and symbolic reasoning, built using the Cognitive Systems Toolkit (CST) and tested in the CoppeliaSim simulation environment. Inspired by Peircean abductive logic, the system explores how artificial agents can construct internal models of causality to perform under uncertain and dynamic conditions.
The system consists of four main layers:
- Interface with CoppeliaSim via RemoteAPI.
- Collect spatial data (position, orientation, rotation) from the ball, striker, and other agents.
- Store recent histories (up to 100 frames) of observed objects.
- Enable temporal tracking for motion analysis.
Ball LearnerandStriker Learner: Linear Data Classifiers trained on historical trajectories.- Predict future positions based on past movements and context.
- Aggregates semantic predictions into behavioral abstractions called "habits".
- Infers causal relations to trigger proactive goalkeeper behaviors.
Each component utilizes CST's idea representation, a flexible symbolic structure supporting cognitive integration and reasoning.
- Based on RoboCup SPL specifications.
- Field dimensions: 9000mm x 6000mm.
- Simulated in CoppeliaSim with physics-based modeling.
- Striker (NAO robot): Locates and kicks the ball.
- Goalkeeper (NAO robot): Tracks the ball laterally and infers striker intent through semantic and causal modules.
- Passive Players: Randomly placed to simulate occlusions and increase spatial complexity.
- Goalkeeper is trained in a fixed environment with variable agent positions.
- Physical parameters like ball size and field friction remain constant.
- Five agent instances trained independently for result consistency.
Three key environmental perturbations were introduced:
-
Friction Change
- Increased coefficient from 0.3 to 0.5.
- Tested adaptation to altered field dynamics.
-
Ball Property Change
- Increased ball mass to 0.5 kg and radius to 20 cm.
- Evaluated the robustness of physical modeling.
-
Spatial Complexity Change
- Increased number of passive players from 5 to 7 per team.
- Tested generalization under occlusion and crowding.
- Continuous error reduction observed for both striker and ball predictions.
- Ball predictions were more accurate due to deterministic dynamics.
- Striker trajectory exhibited more variance, reflecting higher model complexity.
| Test Scenario | Ball Prediction Accuracy | Striker Prediction Accuracy | Notes |
|---|---|---|---|
| Baseline (training scene) | High | High | Good generalization |
| Increased Friction | Moderate | High | Overestimates ball range |
| Ball Property Variation | High | High | Slight deviation on curves |
| More Passive Players | High | High | Slight impact on ball turns |
- The system demonstrated strong generalization to novel conditions, particularly in handling agent-driven behaviors (striker).
- Performance decreased slightly for passive dynamics (ball) in changed friction environments.
- Train
- Friction Change
- Ball Property Change
- Spatial Complexity Change
Available here!
To use the models, create a /profine folder in the project root and insert the files in that folder.
- Abductive Reasoning (Peirce): Core of the causal module; the system hypothesizes unobserved causes from observed effects.
- CST Ideas: Serve as the unifying structure for encoding and reasoning about symbolic and perceptual data.
- CoppeliaSim
- Java (Main code)
- Python (for plot scripts)
- Cognitive Systems Toolkit (CST)
Clone the repository and change to nao branch:
git clone https://github.com/leolellisr/causality
cd causality
git checkout nao