Controller Hierarchy
The high-level policy handles navigation intent using LiDAR-like observations, relative goal position, and lookahead path hints. The low-level policy converts those instructions into stable gait and joint-position behavior.
Back to projects
Robotics Project
Quad-Pedyr (Quadruped Robot)
A simulation-first quadruped project centered on hierarchical RL navigation, obstacle-aware locomotion, and the robot platform design that made the system legible end to end.
The main throughline is the autonomy stack: hierarchical PPO control, LiDAR-guided observations, and goal-reaching behavior in simulation. From there, the page transitions into the mechanical design work around actuators, power, and staged hardware validation.
Project Preview
Jump into the navigation stack first.
Use this preview to move straight into the SLAM and locomotion work before transitioning into the underlying robot design track.
jump to slam
offline
live
Section One
RL navigation and SLAM-style path guidance
The navigation stack uses a hierarchical reinforcement learning system where a high-level controller feeds motion commands to a low-level locomotion controller, while RayCaster observations and path hints help the robot navigate toward goals.
Training Setup
Training ran with 4096 parallel simulated Unitree Go2 robots in Isaac Lab, with NVIDIA PhysX 5 handling terrain interaction, collisions, and sensor simulation across flat and rough terrain.
Observed Results
The locomotion controller improved across both flat and rough environments, with especially strong learning on rough terrain and recovery from early foot-slide penalties as the gait matured.
Takeaways
Exteroception materially improved navigation on rough ground, and reward shaping was essential to prevent the agent from exploiting the simulator with poor-but-legal movement shortcuts.
Technical PDF + Showcase
Reference links
- Technical PDF covering controller design, training setup, and results.
- GitHub repository for the quadruped SLAM and navigation work.
- External showcase video for the end-to-end navigation behavior.
- Embedded media showing locomotion, navigation, and one failure case.
Flat walk baseline
Baseline locomotion clip showing the policy stabilizing movement before denser navigation tasks.
Showcase one
Navigation showcase clip highlighting goal-directed traversal behavior.
Showcase two
Additional behavior sample showing the higher-level navigation policy working with locomotion control.
Failure case / blooper
One failure clip is included to show how the navigation stack behaved when the policy broke down under harder conditions.
Section Two
Robot creation and mechanical design
The autonomy work sat alongside a staged design track: validate a single leg in simulation first, then bench test one leg assembly, then scale into a more complete quadruped platform with the right power, sensing, and control backbone.
Design Track
Platform planning behind the navigation work
The design centers on actuator selection, PCB and power planning, and a single-leg validation loop before committing to a full hardware build. That keeps the project grounded in testable milestones instead of a vague end-state robot concept.
Electronics Stack
The electrical stack includes MCU control, actuator options, a 4S-6S LiPo battery path, fused power distribution, DC-DC rails, and CAN / UART / USB routing between Jetson-class compute and motor drivers.
Validation Plan
The initial plan uses Isaac Sim and ROS2 to test a single leg, validate link lengths and joint limits, check actuator torque margins against estimated mass, and trial gait behavior before risking hardware.
Mechanical Constraints
Build details include chamfers on bolt holes, actuator mounting constraints, output plate alignment, and practical fastener / standoff clearances that matter once the CAD becomes hardware.