News
From Tool to Teammate: Opening the Aperture on Human-Robot Teaming
Researchers are using generative artificial intelligence and cutting-edge scene-mapping technology to elevate robots from simple tools to full teammates capable of providing aid in disaster and battlefield scenarios.
Credit: Johns Hopkins APL/Ed Whitman
Imagine a world where medics and soldiers are partnered with robots that can not only assist with complex tasks — such as transporting casualties to safety or maneuvering quickly through cities or rough terrain — but can also advise on problems and adapt to new information without human intervention.
To achieve this future, robots need a series of complex skills. Those skills include the ability to understand natural language and their surrounding environment in real time, to create and execute plans, and to evaluate their progress and replan as needed.
Researchers at the Johns Hopkins Applied Physics Laboratory (APL) in Laurel, Maryland, are using generative artificial intelligence (GenAI) and cutting-edge scene-mapping technology to elevate robots from simple tools to full teammates capable of providing aid in disaster and battlefield scenarios. The team’s work is funded by the Army Research Laboratory through the Army Artificial Intelligence Innovation Institute (A2I2) and through APL’s Independent Research and Development (IRAD) program.
“This research takes advantage of cutting-edge AI technology for a significant step forward in robotics,” said Chris Korpela, the Robotics and Autonomy program manager within APL’s Research and Exploratory Development Mission Area. “Fully autonomous robots will provide new and exciting capabilities in austere environments.”
The Current State of Robotics
Today, it’s possible to buy fairly advanced robots on the open market — although they cost about the same as a luxury car. Out of the box, these robots do not operate on their own and must receive commands via a controller. There is no option to control them using spoken language, as is possible with tasks on cell phones and tablets. Humans must perform the basic tasks of understanding the robots’ surroundings, creating an execution plan for the tasks the robot will perform, evaluating progress and replanning as needed — severely limiting the level to which people can meaningfully team with a robot.
“While they’re useful for many scenarios, current robots are closer to a remote-controlled car than an autonomous vehicle,” said Corban Rivera, a senior AI researcher at APL and principal investigator for this research. “They can be used as a great tool to enable certain operations, but humans can’t take their hands off the wheel, so to speak.”
Opening Robotic Eyes
An international research team with members from APL, Johns Hopkins University, University of Toronto, Université de Montréal, Massachusetts Institute of Technology, Army Research Laboratory and University of Massachusetts Amherst created a technology that enables a more beneficial partnership to enhance robots’ perception and understanding of their surrounding environment. This technology — — enables robots to have a near-human understanding of a 3D environment.
Using the technology, robots create 3D scene graphs that compactly and efficiently represent an environment. Through training on image-caption pairs from large language models (LLMs) and large visual language models, objects in the scene are assigned tags. These tags help robots understand the meanings and uses of objects as well as the relationships between them.
“Many robots in commercial industry are created to work in factories or distribution centers, which are pristine and predictable environments,” Korpela said. “There are very different needs when robots walk through the woods, for example, where there are numerous and unpredictable obstacles in the way, from rocks on the ground to trees in their path.”
ConceptGraphs is open-vocabulary — meaning it is not limited to the language in its training set — which enables humans to give robots instructions in plain language, either in text or voice, rather than through fixed commands. Robots can even support multimodal queries, which combine an image and a question or instruction. For example, when given an image of Michael Jordan and asked to find “something he would play with,” a robot was able to identify and find a basketball in the environment because of its training on image-caption pairs that provide context to images and objects.
“Now, not only can the robot build up a semantic description of the world, but you can query it in natural language,” said Dave Handelman, a senior roboticist at APL and a collaborator on the project. “You don’t have to ask it if it sees a car — you can say, ‘Show me everything with four wheels,’ or ‘Show me everything that can carry me places.’”
In a real-world scenario, this might translate to a medic asking a robot to locate casualties on a battlefield and transport them to safety until the medic can attend to them. The robot would be able to not only identify casualties but also determine what “safety” means and how to achieve it.
While ConceptGraphs resolved several challenges in human-robot teaming, significant obstacles still remained. Scanning and developing an understanding of the environment took a robot several minutes, and as the robot moved through the environment, more time was needed for additional scanning.
Dave Handelman demonstrates ConceptAgent by working with a robot to determine what safety means and how to achieve it.
Credit: Johns Hopkins APL/Ed Whitman
Enter the ConceptAgent
As the APL team continued innovating, they created a multi-scale model to enable real-time robotic perception and understanding of a 3D environment by combining two variants of large visual language models. The robot uses a fast variant, called Contrastive Language-Image Pre-Training (CLIP), to create an approximate 3D map and make headway right away on assigned tasks.
Then, as time permits, the robot uses a slower but more precise variant, based on generative pre-trained transformer (GPT) models, to obtain a more detailed understanding of the surroundings. The interplay of the two large visual language models enables real-time perception and understanding.
Credit: Johns Hopkins APL
To enhance the robot’s ability to create task execution plans, evaluate progress and replan, the researchers created an AI agent framework — an autonomous system named ConceptAgent that uses an LLM as its engine to reason sequentially — and gave it the ability to write and execute its own code.
The agent has the ability to move, speak and interact with the world. A person can give a command to the ConceptAgent as if they were speaking to another person, and the robot can accomplish the task autonomously. When given a problem, the agent is able to reason sequentially and pivot when it hits a roadblock or new information is discovered.
Researchers recently demonstrated this at APL by tasking a robot with identifying an injured animal and relocating it to an emergency sled. Given a 3D environment with two dummy geese — with one wearing a bandage — the robot was able to correctly identify the “injured goose” and accomplish the task.
“We’ve gone from a situation where teaming with robots was very labor-intensive for humans — the language was controlled by humans, the positions of key objects were given by humans and even the execution plan was created by humans — to create a shared understanding of a goal and the world around us to collaborate,” said Rivera. “All the person has to do is give a command in natural language, and the robot can do the rest.”
Currently, the team is exploring the use of ConceptAgent-based robots to serve as extra hands for medics in the field, as well as to serve as sentries for warfighters, but the sky is the limit.
“We envision that, with this technology, robots will be able to help with any mission a warfighter or first responder may undertake,” Korpela said.
Related Work
