Reinforcement Learning 'TRIPPLES' robot’s speed to 5.2 m/s, 'TEACHES' robot bikes to jump

About a year ago, Boston Dynamics released a research version of its Spot quadruped robot with a low-level API that allows direct control of its joints. At the time, rumors suggested this API enabled Spot to run significantly faster. These rumors originated from the Robotics and AI (RAI) Institute, formerly known as The AI Institute and the Boston Dynamics AI Institute.

At the ICRA@40 conference in Rotterdam last fall, Marc Raibert confirmed that this was not just speculation. Today, we can reveal how the RAI Institute applies reinforcement learning to enhance Spot’s performance dramatically. The same approach has also enabled an autonomous bicycle that can jump.

See Spot run

A recent video showcases Spot running at 5.2 meters per second (11.6 mph). Out of the box, Spot’s top speed is just 1.6 m/s, meaning RAI’s modifications have more than tripled its speed.

https://youtu.be/z478hmH5OME

Spot’s running style may seem unusual because it differs from that of real dogs. "The gait is not biological, but the robot isn’t biological," explains Farbod Farshidian, a roboticist at RAI. Spot’s actuators and kinematics differ from muscle-driven movement, making a dog’s natural running gait unsuitable.

Farshidian describes Spot’s motion as a modified trotting gait with an added flight phase—when all four feet leave the ground. This phase is essential for maintaining high speeds and emerged as a "discovered behavior." Instead of being explicitly programmed to run, Spot learned to move as fast as possible through reinforcement learning.

Reinforcement Learning vs. Model Predictive Control

Boston Dynamics' standard Spot controller relies on model predictive control (MPC). This approach creates a software model of the robot's dynamics, solving optimization problems in real time. While reliable, MPC is limited by the accuracy of its initial model and computational constraints.

Reinforcement learning (RL), in contrast, learns offline. It can use highly detailed models and train a control policy in simulation before deploying it efficiently on the robot. In Spot’s case, modeling every actuator in real-time would be infeasible. Instead, traditional model-based control systems use simplified assumptions to ensure safety and reliability.

Farshidian highlights a surprising discovery: Spot’s speed limit wasn’t due to actuator constraints but rather insufficient battery power. "I thought we were going to hit the actuator limits first," he admits. If they had access to battery voltage data, they could have pushed Spot even faster.

RAI’s RL techniques aren’t just about speed. They could also optimize Spot for efficiency, extending battery life, or reducing noise for indoor environments. By incorporating real-world data into simulations, they can refine robotic capabilities across various applications.

Ultra mobility vehicle: Teaching robot bikes To jump

Reinforcement learning isn’t only useful for improving performance—it also enhances reliability. The RAI Institute developed a unique robot, the Ultra Mobility Vehicle (UMV), a small, self-balancing bicycle trained for parkour using the same RL framework as Spot.

https://www.youtube.com/watch?v=ATWR25xGF74

Unlike traditional bikes, the UMV lacks independent stabilization systems like gyroscopes. Instead, it balances using precise actuator movements. "We’re demonstrating two things," says Marco Hutter, RAI’s Zurich office director. "First, RL enhances the UMV’s driving stability in diverse conditions. Second, understanding dynamic capabilities lets us achieve feats like jumping onto a table higher than the robot itself."

According to Hutter, RL excels in discovering new behaviors and making them robust under unpredictable conditions. Even seemingly simple tasks, like riding backward, are extremely challenging. "Going backward is highly unstable," he explains. "Using classical MPC, we couldn’t do it reliably—especially on rough terrain."

RAI’s research is redefining what robots can achieve. From running robot dogs to jumping bicycles, reinforcement learning is unlocking new levels of agility and performance in robotics.

Comments

Post a Comment

Popular posts from this blog

5 Portable USB Gadgets That Make Everyday Life More Convenient

Image
Modern lifestyles demand flexibility. Whether you're working remotely, traveling frequently, or simply trying to reduce clutter at home, traditional appliances can sometimes feel restrictive. They take up space, depend heavily on wall sockets, and often lack the portability that many people now expect from everyday tools. This post contains affiliate links. If you buy through them, I may earn a commission at no extra cost to you. Thank you! That’s where USB-powered gadgets step in. These compact yet capable devices are designed to operate using standard USB connections, making them compatible with laptops, power banks, and car chargers. Instead of being tied to a single location, you gain the freedom to use them virtually anywhere. Today’s USB-powered appliances are no longer simple accessories . They have evolved into reliable alternatives that can replace larger traditional devices in many situations. Below are five impressive USB-powered gadgets—presented in a completely dif...
Read more

Nothing’s Design-First Vision: How Carl Pei Is Reimagining Consumer Technology

Image
Over the past few weeks, London-based tech company Nothing has introduced a trio of new devices: Phone (4a), Phone (4a) Pro, and Headphone (a). Each product reflects the company’s long-standing promise—deliver thoughtful design and solid performance without pushing consumers into premium price territory. Yet according to co-founder Carl Pei, affordability alone does not define the brand’s appeal. In fact, price is rarely the main reason people choose Nothing devices. This distinction becomes particularly important at a time when the cost of producing electronics continues to rise due to supply chain challenges and memory shortages. Despite these industry pressures, Nothing appears to be positioning itself carefully, relying less on aggressive pricing strategies and more on its distinctive philosophy. Why People Choose Nothing — And It’s Not Just About Price During a conversation at the company’s London headquarters, Carl Pei conveyed a calm but deliberate presence. His understated...
Read more

TSMC pushes chip innovation forward without relying on costly new ASML machines

Image
The global semiconductor industry is often defined by rapid technological leaps and expensive equipment upgrades. Yet recent announcements from Taiwan Semiconductor Manufacturing Company (TSMC) suggest that innovation doesn’t always require buying the newest —and most expensive—tools. Instead, the company is focusing on improving the performance of existing technologies while introducing new methods to enhance computing power. Hot Shot’s Secret Gasoline Extreme 32 oz – Treats 50 Gallons – Concentrated Fuel Additive for Gasoline Engines XEM CHI TIẾT STA-BIL Fast Fix - Small Engine Treatment, Cleans Carburetors and Injectors, Fixes Rough Running Engines, Eliminates Water XEM CHI TIẾT This strategic shift could influence how the entire chipmaking sector evolves, particularly as demand for artificial intelligence (AI), data centers, and high-performance computing continues to rise. Rather than chasing every new piece of hardware, TSMC is demonstrating how thoughtful engineering and re...
Read more