Meet Roadrunner: RAI Institute's Bipedal Wheeled Robot That Thinks in Two Modes

What happens when you give a bipedal robot wheels instead of feet — and then let it choose how to orient them? RAI Institute's new Roadrunner answers that question with a compact, 15 kg platform that can switch between two fundamentally different wheel configurations on the fly, all while relying on a single neural network policy trained purely through reinforcement learning in simulation.

Two Configurations, One Brain

Most wheeled bipeds pick a lane: either both wheels sit side-by-side (like a Segway) or they roll in-line (like a bicycle). Roadrunner does both. Its legs can position the wheels in a side-by-side configuration for broad lateral stability or shift into an in-line configuration to navigate narrow corridors and squeeze through tight spaces. This isn't a manual mode switch — the robot transitions dynamically depending on terrain demands and operator commands.

What makes the control story compelling is the scale of the challenge: side-by-side and in-line represent fundamentally different balance problems. A side-by-side stance offers natural lateral stability but requires pitch control, while in-line demands constant active balance in the roll axis. Most teams would train separate policies. RAI Institute trained one — and validated it zero-shot, meaning the sim-trained policy transferred directly to hardware without any real-world fine-tuning.

Symmetric Legs: Knees Forward or Backward

Roadrunner's legs are fully symmetric in design, which unlocks a trick that most humanoids can't pull off: the knees can point either forward or backward. This isn't an aesthetic choice. When facing an obstacle or navigating uneven ground, flipping the knee direction changes the robot's effective reach and ground clearance. A knee pointing forward lowers the center of mass and increases pushing force; one pointing backward extends reach and helps step over or around obstacles. Combined with the wheel-configuration switching, Roadrunner has a much richer vocabulary of motion than its modest 15 kg frame suggests.

Reinforcement Learning All the Way Down

The control system is built on a single reinforcement learning policy that was trained entirely in simulation. This sim-to-real transfer approach — made famous by systems like OpenAI's Dactyl and Boston Dynamics' internal RL experiments — is now increasingly the standard for agile locomotion. RAI Institute's contribution here is showing that a multi-modal platform, one with mechanically reconfigurable geometry, can also benefit from this paradigm. The policy doesn't just learn to balance; it learns when and how to reconfigure.

Zero-shot transfer is the hardest test of sim-to-real robustness. There's no teacher in the real world correcting the robot's mistakes — the policy either generalizes or it falls over. Roadrunner's successful zero-shot deployment suggests RAI Institute has invested heavily in domain randomization and accurate actuator modeling during training.

Where Roadrunner Fits

Wheeled bipeds occupy an interesting niche: they're faster and more energy-efficient than legged walkers on flat ground, but they retain the leg-height advantage for dealing with steps, curbs, and cluttered environments. Platforms like Agility Robotics' early Cassie experiments and Boston Dynamics' Handle showed the potential of the form factor years ago. Roadrunner refines the concept with a focus on mechanical configurability — the idea that adaptability shouldn't come only from software but should be embedded in the hardware architecture itself.

At 15 kg, Roadrunner is lightweight enough to be practical in indoor environments — warehouses, offices, research labs — without the infrastructure overhead of a full humanoid. Its symmetric leg design also simplifies manufacturing and spare-parts logistics, a consideration that matters once you move beyond prototype counts.

What's Next for RAI Institute

RAI Institute has been quietly building a portfolio of locomotion research, and Roadrunner represents a significant step toward deployable hardware. The team has not yet announced commercial timelines, but the zero-shot sim-to-real result and the mechanical reconfigurability suggest the platform is far enough along to serve as a serious research and eventual deployment vehicle. Expect follow-up work on whole-body manipulation — once a wheeled biped can handle both navigation modes, adding arm-based tasks is the natural next frontier.

"A single RL control policy handles both wheel configurations zero-shot — no real-world fine-tuning required."

— RAI Institute, LinkedIn announcement