The Newton Capstone

An open-source quadruped robotics platform built to open-source and lower the barrier of entry of legged locomotion designed for students, small research groups, and makers. Working with a team of 7 engineering students, we integrated mechanical design, embedded systems, control theory, and reinforcement learning, to produce a realistic, low-cost robotics platform.

We even have a blog with some details of our process here and our final report here.

Final Project

Newton is an 8-DoF quadruped with a fully open hardware and software stack. Although originally aimed for 12 DoF, technical and time constraints pushed us toward an optimized 8-DoF design, that can still walk with some agility, and runs a complete sim → training → deployment pipeline within 15 minutes.

Overall, we ended with a mobile robotics platform that has:

Real walking gaits on flat terrain
Reinforcement learning–trained policies deployed on real hardware
Onboard perception and mapping via Isaac ROS
Modular legs and actuators for quick experimentation
Open-source design files, firmware & software

Prep Project

Before diving into building Newton, our robot dog, we started building TWIP, a Two-Wheeled Inverted Pendulum robot, designed and ideated by gym2real, a Capstone team, in the University of British Columbia. We wanted a straightforward way to learn about the Reinforcement Learning (RL) pipeline with a robot that we can easily, and cheaply, build.

Technical Highlights

Mechanics & Actuation

Belt-driven 9:1 dual-stage transmission
SLS-printed gears & transmission
FDM-printed PLA/PETG chassis
Lightweight (~3.4 kg) and low cost (~4,000 USD)

Electronics

ODrive S1 controllers with 340KV BLDC motors
6S LiPo battery (~1-hour runtime)
Jetson Orin Nano for compute and perception
CAN bus connectivity across controllers & Jetson Orin Nano

Control & Firmware

Hierarchical control layers: 8 kHz motor loop, 50 Hz joint control, 30 Hz perception
ROS2-based middleware within the Jetson Orin Nano
Real-time logging, fault detection, and safety handling

Simulation & Learning

RL training using Isaac Sim and later, with more success, Genesis
Domain randomization (friction, mass, terrain) for a smaller gap transfer
Modular simulation components (environment, task, agent)

Perception & Navigation

Visual SLAM (Isaac ROS)
3D mapping with Nvblox
Navigation using ROS2’s nav2 stack (simulated)

Challenges & Trade-offs

If we learnt anything working on Newton, is that pivots are essential.

Early in the project, we pivoted from ESP32-based controllers to ODrive for better communication performance. Similarly, we moved from Isaac Sim to Genesis to stabilize training.

Mechanical trade-offs and design decisions, like pulley design, belt tension, and encoder alignment required multiple design cycles.

Finally, despite many attempts at using Isaac Sim as our RL simulator (and over 7 months invested into the platform), we pivoted to Genesis since it provided a Python API that was easier to work with and overall better results (both in simulation and on-robot).

Results & Impact

We didn’t achieve everything we planned to do, but we achieved so much more than we expected.

Newton already fulfills its mission of having an affordable, and open quadruped robotics platform and we wish that the project continues on.