RAI calls the bicycle robot an ultra-mobility vehicle (UMV). It can reach a height of 3 feet and can jump from the floor onto a platform.
The contributions of a Georgia Tech Ph.D. student helped make these feats possible through a robot control policy he developed.
Jeonghwan Kim, who is pursuing a Ph.D. in robotics under the advisement of Associate Professor Sehoon Ha, spent two semesters interning at RAI. His task was to design a policy to teach the UMV to land after a flip.
The result was iterative motion imitation (IMI), a novel method that imitates flip trajectories generated from prior examples. Kim said the robot bases its flip on a demonstration, and human engineers reconstruct and refine the flip path through simulation to fill in the gaps.
“To guide the robot to flip, we started with an imperfect trajectory generated by a motor-based controller and then ran simulations,” Kim said. “It’s an unstable trajectory, but we use it as a guide to train a single policy that can track it as it lands and tries to balance itself.”
Kim interned under the supervision of Shamel Fahmi, a research scientist at the RAI Institute. RAI has been developing the UMV for nearly three years.
“We wanted to work on a different robot morphology that wasn’t legs or legs with wheels,” Fahmi said. “That’s when we thought of working with bikes.
“We want to merge the athleticism of (Boston Dynamics’) Atlas with the mobility of a bike. We wanted a robot that can go anywhere, do parkour, and acrobatics.”
Fahmi said that before Kim arrived, the research team had trouble getting the UMV to land consistently without breaking or falling.
The UMV has two joints — an upper and a lower. The upper joint contains the motors and pulls the lower joint along as it propels into the air. The problem is getting the lighter lower joint to absorb the impact of landing without being crushed by the heavier upper joint.
“That’s what brings reinforcement learning into the equation,” Fahmi said. “We teach the robot to minimize its impact on the ground to land gracefully.”
Fahmi said that Kim proved the imitation examples the robot learns from don’t have to be perfect. The process takes some time, but all it needs is a rough idea to get started.
“You can have an imperfect sketch and then constantly refine it,” Fahmi said. “The first time, it’s not going to go well.
“We don’t care about torque or power limits as long as it does the motion. Then we’ll have a slightly better reference, repeat it, and imitate it again. In every iteration, we can add more parameters.”
Kim said he felt the pressure of time constraints during his two semesters with RAI as he worked to achieve consistent, successful landings. Even though he had multiple UMVs to experiment with, they broke down dozens of times. Each time one broke, a hardware team at RAI had to repair it.
“There was a lot of pressure to not only get this working before my internship ended, but also knowing there are costs behind every failed attempt, and every time the robot breaks, it takes time to repair it,” Kim said.
“It took almost five months for it to land without breaking. Then we needed two more months for it to stay balanced after the landing. It requires a lot of engineering effort to achieve a robust control policy for a safe flip.”
By the time Kim left RAI, the IMI policy had achieved consistent, seamless landings.
“The jump right now is what we call the visitor demo,” Fahmi said. “If there are guests coming over to see it, we want to show them something that is extremely impressive, but also, more importantly, extremely reliable. It never fails.
“It was only possible because of the huge effort we put into designing, maintaining, and continuously improving the robot.”
Kim authored a paper on his framework and will present it at this week’s International Conference on Robotics and Automation (ICRA) in Vienna.
For more information about the UMV project, please visit the RAI blog or watch their video on YouTube.
]]>Jeonghwan Kim, who is pursuing a Ph.D. in robotics under the advisement of Associate Professor Sehoon Ha, spent two semesters interning at RAI. His task was to design a policy to teach the UMV to land after a flip.
]]>The new technology is also set to revolutionize the scale of policy training data collection, which is essential to advancing robotic capabilities and meeting growing production demand.
COBALT is a mobile app that turns smartphones into controllers for robot arms. With a secure Wi-Fi connection to a server, users can move their phones in any direction, and the robot arm will mirror the motion — from anywhere in the world.
Ayush Agarwal, a Ph.D. student in Georgia Tech’s School of Interactive Computing who leads a research team developing COBALT, said it works like the games people play on smartphones. Users can press a button to have the arm grasp an object, move it, and release it with another button.
Agarwal conducted several user studies with participants in nine countries who remotely operated robot arms inside Georgia Tech’s People, AI & Robotics (PAIR) Lab. The lab is directed by Assistant Professor Animesh Garg, who advises Agarwal.
“We built an entire distribution system for remote teleoperation scaled to where we had people from Indonesia, India, and Pakistan operating for us,” Agarwal said. “They were novice operators who had never done it before. By collecting data from these new users, we showed that we can train policies to automate certain tasks.”
Garg envisions a world where data collection for policy training is done through crowdsourcing. He began working toward this goal 10 years ago as a postdoc at Stanford University, when he developed RoboTurk, an earlier version of COBALT.
“There is a large-scale data collection requirement for mass robot production to be possible, and it will not be solved purely through simulation,” Garg said.
“Our idea was, what if we could get almost every person on the planet to be a passive source for data collection? There are almost five billion people who have smartphones and know how to use them.”
Another major implication of COBALT could be expanded access to CS and robotics education.
Students can learn to operate a robot remotely in any classroom. In fact, Garg and his lab recently hosted students from Midtown High School in Atlanta to demonstrate COBALT and let them control robot arms from a phone.
Garg also sees the possibility of a “gig economy” in which people pay remote operators to control assistive robots in their homes and complete household chores for them.
“It could be Uber for robots,” he said. “People who want to log onto the platform can do so at their convenience and for as long as they want.”
Companies with robot-dependent labor tasks could also use the platform to enable human oversight.
“If I deploy a robot in a factory that achieves high autonomy for most tasks, but there are still times it needs help, a human could operate the robot from anywhere in the world,” Garg said.
Agarwal’s studies showed that people prefer to interact with and control a robot using a smartphone rather than virtual reality (VR) headsets, controllers, keyboards, mice, or other devices.
“The phone is a more intuitive interface and can provide data quality that’s on par with other commonly used devices,” he said.
Agarwal also said there is minimal latency in the video feed sent back to operators on the other side of the world. That’s because the amount of data being processed is small.
The data is carried over Web Real-Time Communication (WebRTC), the same technology used by many streaming services and web conferencing platforms such as Zoom and Google Meet.
“There’s a connection from your phone to the teleoperation server, which is connected to the robots,” Agarwal said.
“Then there’s another connection from the teleoperation server back to the user, which allows for a video stream. We need low latency on both because you don’t want the user to move their phone and wait 10 seconds to see the visual feed.”
Agarwal is the co-lead author of a paper on COBALT that is being presented at the IEEE International Conference on Robotics and Automation this week in Vienna. He said the paper stands out because it has moved from theory to the implementation of an entire distribution network.
“The real novelty of our paper is the systems that we build around it to actually support the scaling of remote operation and data collection at a global level,” he said.
]]>Indoor farms, also known as vertical farms, are popular among agricultural researchers and are expanding across the agricultural industry. Some benefits they have over outdoor farms include:
Additionally, some studies indicate that indoor farms produce more nutritious food for urban communities.
However, these farms are often inaccessible to birds, bees, and other natural pollinators, leaving the pollination process to humans. The tedious process must be completed by hand for each flower to ensure the indoor crop flourishes.
Ai-Ping Hu, a principal research engineer at the Georgia Tech Research Institute (GTRI), has spent years exploring methods to efficiently pollinate flowering plants and food crops in indoor farms to find a way to efficiently pollinate flower plants and food crops in indoor farms.
Hu, Assistant Professor Shreyas Kousik of the George W. Woodruff School of Mechanical Engineering, and a rotating group of student interns have developed a robot prototype that may be up to the task.
The robot can efficiently pollinate plants that have both male and female reproductive parts. These plants only require pollen to be transferred from one part to the other rather than externally from another flower.
Natural pollinators perform this task outdoors, but Hu said indoor farmers often use a paintbrush or electric tootbrush to ensure these flowers are pollinated.
An early challenge the research team addressed was teaching the robot to identify the “pose” of each flower. Pose refers to a flower’s orientation, shape, and symmetry. Knowing these details ensures precise delivery of the pollen to maximize reproductive success.
“It’s crucial to know exactly which way the flowers are facing,” Hu said.
“You want to approach the flower from the front because that’s where all the biological structures are. Knowing the pose tells you where the stem is. Our device grasps the stem and shakes it to dislodge the pollen.
“Every flower is going to have its own pose, and you need to know what that is within at least 10 degrees.”
Harsh Muriki is a robotics master’s student at Georgia Tech’s School of Interactive Computing, who used computer vision to solve the pose problem while interning for Hu and GTRI.
Muriki attached a camera to a FarmBot to capture images of strawberry plants from dozens of angles in a small garden in front of Georgia Tech’s Food Processing Technology Building. The FarmBot is an XYZ-axis robot that waters and sprays pesticides on outdoor gardens, though it is not capable of pollination.
“We reconstruct the images of the flower into a 3D model and use a technique that converts the 3D model into multiple 2D images with depth information,” Muriki said. “This enables us to send them to object detectors.”
Muriki said he used a real-time object detection system called YOLO (You Only Look Once) to classify objects. YOLO is known for identifying and classifying objects in a single pass.
Ved Sengupta, a computer engineering major who interned with Muriki, fine-tuned the algorithms that converted 3D images into 2D.
“This was a crucial part of making robot pollination possible,” Sengupta said. “There is a big gap between 3D and 2D image processing.
“There’s not a lot of data on the internet for 3D object detection, but there’s a ton for 2D. We were able to get great results from the converted images, and I think any sector of technology can take advantage of that.”
Sengupta, Muriki, and Hu co-authored a paper about their work that was accepted to the 2025 International Conference on Robotics and Automation (ICRA) in Atlanta.
The pollination robot, built in Kousik’s Safe Robotics Lab, is now in the prototype phase.
Hu said the robot can do more than pollinate. It can also analyze each flower to determine how well it was pollinated and whether the chances for reproduction are high.
“It has an additional capability of microscopic inspection,” Hu said. “It’s the first device we know of that provides visual feedback on how well a flower was pollinated.”
For more information about the robot, visit the Safe Robotics Lab project page.
]]>A Georgia Tech research led by Ai-Ping Hu and Shreyas Kousik team is working to solve that. A robot they've developed can efficiently pollinate plants that have both male and female reproductive parts. These plants only require pollen to be transferred from one part to the other rather than externally from another flower.
]]>In these contests, a handler directs a trained dog with whistle signals to guide a small group of sheep across a field and sometimes split the flock cleanly into two groups. But sheep do not always cooperate.
Researchers at the Georgia Institute of Technology studied how handler–dog teams manage these unpredictable flocks in sheepdog trials and found principles that extend beyond livestock herding.
In a study published in Science Advances as the cover feature, the researchers applied those insights to computer simulations showing how similar strategies could improve the control of robot swarms, autonomous vehicles, AI agents, and other networked systems where many machines must coordinate their actions despite uncertain conditions.
Group Movement Dynamics
“Birds, bugs, fish, sheep, and many other organisms move in groups because it benefits individuals, including protection from predators,” said Saad Bhamla, an associate professor in Georgia Tech’s School of Chemical and Biomolecular Engineering. “The puzzle is that the ‘group’ is not a single organism. It is built from many individuals, each making local, imperfect decisions.”
When a predator threatens a herd of sheep, individuals near the edge often move toward the center to reduce their own risk, Bhamla explained. “This is ‘selfish herd’ behavior,” he said. “Shepherds exploit that instinct using trained dogs.”
From examining hours of contest footage, the researchers found that controlling small groups of sheep can be harder than managing large ones. A larger group, with more sheep protected in the center, may behave more coherently than a small group as the animals constantly shift between two instincts: “follow the group” and “flee the dog.”
“That switching behavior makes the group unpredictable,” said Tuhin Chakrabortty, a former postdoctoral researcher in the Bhamla Lab who co-led the study.
Looking closely at how dogs and their handlers guide small groups, the researchers found that unpredictability in the flock’s behavior does not always make control harder. “Under the right conditions, that ‘noisy’ behavior might actually be a benefit,” Bhamla said.
Successful Sheep Herding
Sheepdog handlers categorize sheep by how strongly they respond to a dog’s threatening pressure. Some very responsive sheep might panic under too much pressure, while others might ignore mild pressure and require stronger positioning by the dog.
The researchers observed that successful control often followed a two-step pattern. First, the dog subtly influenced the sheep’s orientation while the animals were mostly standing still. Once the flock was aligned in the desired direction, the dog increased pressure to trigger movement. The timing of those actions was critical, because alignment within a small group could disappear quickly as individuals switched between instincts.
“In our simulations, increasing pressure makes the flock reach the desired orientation faster, but how long the flock stays aligned is set mainly by noise,” Chakrabortty said. “In essence, dogs can steer the direction, but they can’t hold that decision indefinitely, so timing matters.”
Developing Computer Models
To understand the broader implications of that behavior, the team developed computer models that captured how sheep respond both to the dog and to one another. The models allowed the researchers to test different strategies for guiding groups whose members make independent decisions under uncertainty.
They then applied those ideas to simulations of robotic swarms. Engineers often design such systems so that each robot blends signals from all nearby robots before deciding how to move. While that approach works well when signals are clear, it can break down when information is noisy or conflicting, Bhamla explained.
To explain why that switching strategy can work under noisy conditions, the researchers used an analogy of a smoke-filled room where only one person can see the exit, and no one knows who that person is. If everyone polls everyone else and averages the guesses, the one correct signal can get diluted by many noisy ones.
“That’s the counterintuitive part. When only one person has the right information, averaging can wash out the signal. But if you follow one person at a time, and keep switching who that is, the right information can spread through the crowd,” Bhamla said.
Building on that idea, the researchers tested a strategy inspired by the switching behavior they observed in sheep. In the simulations, each robot paid attention to just one source at a time (either a guiding signal or a neighboring robot) and switched that source from one step to the next.
Under noisy conditions, this switching strategy required less effort to keep the group moving along a desired path than either averaging-based strategies or fixed leader-follower strategies.
The researchers call their approach the Indecisive Swarm Algorithm. The name reflects a counterintuitive insight: allowing influence to shift among individuals over time can make groups easier to guide when conditions are uncertain.
“Our findings suggest that the same dynamics that make small animal groups unpredictable may also offer new ways to control complex engineered systems,” Bhamla said.
CITATION: Tuhin Chakrabortty and Saad Bhamla, “Controlling noisy herds: Temporal network restructuring improves control of indecisive collectives,” Science Advances, 2026
This research was funded in part by Schmidt Sciences as part of a Schmidt Polymath grant to Saad Bhamla.