Joint Field Characterisation of Onboard Perception, Mapping and Manipulation on a Sub-13 kg Rocker-Bogie Rover
This paper presents the joint outdoor characterization of a sub-13 kg rocker-bogie rover running concurrent perception, mapping, and manipulation tasks on a single embedded GPU, revealing that a fully integrated autonomy stack achieves high performance while drawing significantly more power than the sum of its isolated subsystems would predict.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Robots that can see, map their surroundings, and pick up objects are no longer just science fiction; they are becoming a reality for exploring places humans cannot easily reach. To work in the wild, these machines need to be small enough to carry their own power and computers, yet powerful enough to make decisions on the fly. The challenge lies in fitting all these demanding tasks—recognizing what is in front of them, building a mental map of the terrain, and controlling an arm to grab things—into a single, lightweight package. When engineers design these systems, they often test each part separately in a lab, assuming that if the camera works well alone and the motor works well alone, the whole robot will work well together. However, in the real world, these parts compete for the same battery and the same computer chip. Understanding how they actually behave when running all at once, especially on a small vehicle moving over grass, is crucial for knowing how long a robot can truly last and how safely it can operate.
A team of researchers at the Vellore Institute of Technology in India set out to test this exact scenario. They built a small, six-wheeled rover weighing just under 13 kilograms, roughly the size of a large suitcase but with a suspension system designed to keep its body level even on rough ground. This vehicle was equipped with a camera, a laser scanner that measures distance, and a six-jointed robotic arm. All the "brain" work was done by a single, compact computer module the size of a credit card, which had to handle everything from spotting a target to steering the wheels and moving the arm. The researchers took this machine outdoors onto a grassy field to see how it performed when every system was running simultaneously, rather than in isolation.
The results revealed a complex picture of success and limitation. The robot's vision system, tasked with finding a specific 12-centimeter cylindrical object, was quite effective. Even while the robot was also mapping its surroundings and planning its path, the camera identified the target correctly in nearly 90 percent of the test images. The robot moved at a steady pace of about 14.5 images per second, which was fast enough to keep up with its laser scanner. When it came to navigation, the robot built a map of a 10-meter by 10-meter grassy area with a positional accuracy of about 11 centimeters. However, this result was achieved only because the test plot was surrounded by traffic cones and fencing to provide the vertical structures the laser scanner needs; the paper notes that this accuracy figure applies to this bounded plot and says nothing about performance on open grass. This level of precision allowed the robot to detect obstacles and re-plan its route in just 220 milliseconds, a speed fast enough to keep the vehicle safe at a top speed of roughly 0.34 meters per second.
The physical manipulation task was also largely successful. Over 30 attempts to find, approach, and pick up the cylinder, the robot succeeded 23 times, a success rate of about 77 percent. However, the paper emphasizes that this point estimate comes with a wide 95% confidence interval of 0.59 to 0.88, meaning the true success rate could be significantly lower or higher, and the small sample size prevents a statistically reliable separation of failure modes. Most of the failures happened not because the arm was too weak or the gripper too clumsy, but because the robot stopped in the wrong spot. The arm was mounted in a fixed position, so if the wheels stopped even a few centimeters off-target, the arm could not reach the object. This highlighted that the robot's ability to grab things was limited more by how accurately it could park itself than by the strength of its mechanical arm.
Perhaps the most significant finding concerned the robot's energy consumption. Engineers often try to predict how long a robot will last by adding up the power used by its camera, its motors, and its arm when each is tested separately. In this study, when the researchers added up the power usage of the robot's individual systems in isolation, they predicted the robot would use about 22 watts of power during a mixed task. However, when they ran the robot with all systems active at the same time, it actually drew 38.5 watts. This means the real-world power demand was nearly double what the separate tests suggested. The extra power was consumed because the systems were interfering with each other, creating a load that simple addition could not predict. Based on this actual measured power draw, the researchers calculated that the robot could operate for about 5.3 hours on a single battery charge. It is important to note that this endurance figure is a derived calculation based on an assumed battery depth of discharge (90%) rather than a measured runtime to battery cutoff, meaning the actual time could vary depending on that assumption.
The study also examined the robot's ability to communicate with its human operator. The team used a standard radio link to send status updates, but they found that over distances of 500 meters, the connection became unreliable, losing more than one in every nine messages. Because of this, the researchers designed the robot so that the emergency stop button was a physical switch on the vehicle itself, completely independent of the radio. This ensured that even if the connection failed, a human could still stop the robot by pressing the button, a safety measure that proved essential given the limitations of the wireless signal in a grassy environment.
Ultimately, this research demonstrates that building a capable, small-scale robot requires more than just assembling good parts. It requires understanding how those parts interact under real-world conditions. The study showed that while the robot could successfully navigate, map, and manipulate objects outdoors, its true limits were defined by the friction of the grass, the precision of its parking, and the hidden energy costs of running everything at once. The findings serve as a reminder that for small robots, the whole is often more demanding than the sum of its parts, and that accurate field testing is the only way to know what a machine can truly do.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.