Quadruped Walking Robot

The mechanical design began with concept sketches evaluating different locomotion approaches before settling on a quadruped configuration for optimal stability and gait versatility. Using SolidWorks I built a complete parametric model of the robot.

• Four-leg assembly with articulated joints for each leg
• Servo motor mounting brackets and housings
• Custom chassis to accommodate electronics and power systems
• Motion studies to validate range of motion and check for mechanical interference
• Assembly animations to visualize the walking gaits before physical construction

The CAD model was optimized for 3D printing, with careful attention to print orientation, support material requirements and part tolerances.

Before committing to physical hardware, I built a custom simulation environment in MuJoCo (Multi-Joint dynamics with Contact) to develop and test walking gaits.

• An accurate digital twin with matching inertial properties, joint constraints and contact dynamics
• Sinusoidal gait controllers as a baseline for coordinated leg movement
• Machine-learning parameter optimization using Random Search and a Parallel Hill Climber to tune phase offsets, stride length and step frequency

The architecture lets the Raspberry Pi send high-level motion commands, which the servo controller translates into precise joint angles distributed across the daisy chain.

To make the robot accessible and easy to control remotely, I built a Telegram bot interface.

• Real-time command execution through chat messages
• Trigger different gait patterns and behaviors remotely
• Status updates and debugging information sent back to the user
• Convenient testing and demonstration without physical access to the robot

Beyond scripted gaits, I explored learned control policies.

• Built a custom Gym environment wrapping the MuJoCo simulation
• Implemented Proximal Policy Optimization (PPO) for policy training
• Experimented with several reward formulations for stable forward locomotion
• The reward tuning never fully converged, but the process was hands-on learning in RL system design, debugging training instabilities and bridging the sim-to-real gap

• Sim-to-real transfer is hard: motor backlash, surface friction variability and battery voltage sag all create behavior the simulator does not predict
• Iteration is essential: multiple design-test-refine cycles were needed for both the mechanical and control systems
• Systems integration is complex: power distribution, communication protocols, mechanical constraints and control algorithms all constrain each other

The project culminated in a functioning quadruped capable of autonomous walking, with the flexibility to execute different gaits and respond to remote commands.