Robotics simulation and machine learning · Active development
MechaSim
An experimental Windows robotics simulation and training application for constructing mechanisms, testing control strategies, and exploring learned locomotion.
Experimental robotics workspace
MechaSim is a Windows-based application for constructing or importing robotic mechanisms, configuring physical properties, defining training objectives, and experimenting with robot control. The longer-term research goal is to investigate whether controllers learned in simulation can eventually transfer to physical robots. That sim-to-real goal has not yet been validated.
The application was formerly called ForgeSim; MechaSim is its public name.
Documented capabilities
The application guide covers geometry import, mechanism setup, rigid bodies, joints, motors and drives, articulated robots, physics-based testing, authored motion sequences, inverse kinematics, and neural movement training. Training workflows include configurable objectives, staged progression, PID and feedforward gain experiments, manual joint control, compatible physical game-controller input, and parallel simulations. Projects and trained controllers can be retained. Experimental AI/MCP integration and board-and-motor export functionality are also documented.
Hexapod and wheeled-biped experiments
In internal experiments, a simulated six-legged robot learned balanced walking, destination-directed movement, some obstacle navigation, and controller-responsive motion. A wheeled biped progressed from being unable to stand to balancing and navigating toward targets, with some Xbox controller responsiveness.
Owen reports that useful balancing and locomotion behaviour emerged after approximately 10–16 hours of training using as many as ten parallel simulation instances. These observations come from internal experimentation and are not independently verified benchmark results.
Current limitations
Learned behaviour remains imperfect. Some target directions still cause navigation or reorientation problems, and no physical sim-to-real transfer has been demonstrated. There is not yet independently verified benchmark data. MechaSim should therefore be understood as an experimental learning and development environment, not a proven real-world robot-control platform.
Next investigations
Future work includes clearer measured comparisons across training stages, further navigation and reorientation improvements, richer experiment documentation, and cautious investigation of hardware integration and controller export. Physical deployment will require separate validation and safety-conscious testing.
AI-assisted development disclosure
Owen conceived MechaSim, defines its product requirements and direction, researches control approaches, tests functionality, evaluates simulated behaviour, and iteratively directs development. A significant portion of the source-code implementation is produced with AI coding tools such as Codex. This project demonstrates requirements definition, technical investigation, test-driven iteration, and AI-assisted product development; it does not imply that Owen personally authored every simulation or neural-network algorithm.
Development, step by step.
Mechanism and control workspace
Established workflows for imported geometry, articulated mechanisms, manual control, motion authoring, and physics testing.
Hexapod locomotion
Observed balanced walking, destination-directed movement, limited obstacle navigation, and controller response in simulation.
Wheeled-biped balance
Progressed from failure to stand toward balancing and target-directed simulated movement.
Validation remains open
Current work focuses on imperfect behaviours, reproducible measurements, and the unvalidated sim-to-real question.