Each agent broadcasts its trajectory so that the other drones know where it intends to go. The agents then consider each other’s trajectories as they optimize their own to ensure they do not collide.

But when the team tested the system on real drones, they found that if the drone didn’t have up-to-date information about its partner’s trajectory, it could inadvertently choose a path that led to a collision. The researchers modified their system and are now rolling out Robust MADER, a multi-agent trajectory planner that generates collision-free trajectories even when communication between agents is delayed.

“MADER works well in simulations, but it has not been tested in hardware. Therefore, we built a batch of drones and started flying them. The drones need to talk to each other to share trajectories, but once you start flying, you quickly realize that there are always communication delays that introduce some glitches,” said Kota Kondo, a graduate student in aeronautics and astronautics.

Making multiple drones work together

When multiple drones are working together in the same airspace, they run the risk of colliding. But now researchers at AeroAstro have created a trajectory planning system that allows drones in the same airspace to always choose a safe path forward. Credit: Courtesy of the researchers

The algorithm includes a delay-checking step, during which the drone waits a certain amount of time before committing to a new, optimized trajectory. If it receives additional trajectory information from other drones during the delay, it may abandon its new trajectory and restart the optimization process.

When Kondo and his collaborators tested Robust MADER in flight experiments with simulated and real drones, it achieved a 100 percent success rate in generating collision-free trajectories. While the drones’ flight times were somewhat slower than some other methods, no other baseline could guarantee safety.

“If you want to fly safer, you have to be careful, so it makes sense that if you don’t want to collide with an obstacle, it will take longer to get to your destination.” Kondo said, “If you collide with something, it doesn’t really matter how fast you go, because you won’t get to your destination.”

Kondo, along with postdoctoral fellow Jesus Tordesillas, graduate student Parker C. Lusk, MIT undergraduates Reinaldo Figueroa, Juan Rached and Joseph Merkel, and senior author Richard C. Maclaurin Professor of Aeronautics and Astronautics and Information and Decision Systems Laboratory (LIDS) principal investigator and member of the MIT-IBM Watson AI Lab, Jonathan P. How, co-authored the paper. The research will be presented at the International Conference on Robotics and Automation.

Planning trajectories

MADER is an asynchronous, decentralized, multi-agent trajectory planner. This means that each UAV develops its own trajectory, and while all agents must agree on each new trajectory, they do not need to agree at the same time. This makes MADER more scalable than other approaches, since it is difficult for thousands of UAVs to agree on a trajectory at the same time. Because of its decentralized nature, the system will also work better in realistic environments where drones may fly far from a central computer.

With MADER, each drone optimizes a new trajectory using an algorithm that incorporates the trajectories it receives from other agents. By constantly optimizing and broadcasting their new trajectories, drones can avoid collisions.

However, perhaps an agent shared its new trajectory a few seconds ago, but its companion did not receive it immediately due to communication delays. In real-world environments, signals are often delayed by environmental factors such as interference from other devices or stormy weather. Due to this inevitable delay, a drone may inadvertently submit a new trajectory that puts it on a collision course.

Robust MADER completely prevents such collisions because each agent has two available trajectories. It keeps one trajectory that it knows is safe and that it has checked for potential collisions. While flying along the original trajectory, the drone optimizes a new trajectory, but it will not commit to the new trajectory until it has completed the delayed check step.

During the delayed check, the drone spends a fixed amount of time double-checking communications from other agents to see if its new trajectory is safe. If it detects a potential collision, it abandons the new trajectory and starts the optimization process again. The length of the delayed checking period depends on the distance between agents and environmental factors that may impede communication. For example, if the agents are miles apart, then the delayed check period needs to be longer.

Eliminating collisions completely

The researchers tested their new approach by running hundreds of simulations in which they artificially introduced communication delays. In each simulation, Robust MADER was 100% successful in generating collision-free trajectories, while all baselines caused collisions.

The researchers also built six drones and two aerial obstacles and tested Robust MADER in a multi-agent flight environment. they found that while using the original version of MADER in this environment resulted in seven collisions, Robust MADER did not cause a single crash in any of the hardware experiments.

“Until you actually fly the hardware, you don’t know what might cause problems. Because we know there’s a difference between simulation and hardware, we made the algorithm robust so it works in real drones, and it’s very rewarding to see that in practice,” Kondo said.

Using Robust MADER, the drones were able to fly 3.4 meters per second, although their average travel time was slightly longer than some baselines. But no method has been completely collision-free in every experiment.

In the future, Kondo and his collaborators hope to put Robust MADER to the test outdoors, where there are many obstacles and types of noise that can interfere with communication. They also want to equip the drones with vision sensors so that they can detect other agents or obstacles, predict their movements, and incorporate that information into trajectory optimization.

The post MIT releases advanced algorithm to completely prevent drones from colliding in the air appeared first on TechGoing.

The research was led by Jeehwan Kim, an engineering professor who specializes in ultra-thin, high-performance films. The team built on past research to fabricate membranes with red, green and blue sub-pixels on a two-dimensional membrane material.

The films were next peeled from a rigid base layer and then stacked together to form vertical full-color pixels only 4 microns wide.

This is the smallest microLED pixel and the highest pixel density reported in the journal,” said Jeehwan Kim. And stacking colors is just the first step. The team’s in-depth study found that by varying the input voltage, they were able to produce multiple colors in each stacked pixel. But to control 25 million individual LEDs required a sophisticated system. This is one of the team’s future developments.

The post MIT develops vertically stacked MicroLED technology appeared first on TechGoing.

The technology has already been recognized within the industry, and a paper describing it will be accepted at the International Conference on Machine Learning (ICML) in July.

EquBind model can rapidly screen drug-like molecules with 1200 times speedup
Currently, drug development is a long and expensive affair. One of the main reasons is that the cost of developing drugs is very expensive. This cost includes not only billions of dollars of capital investment, but also decades of research time.

And in the process of research and development, 90% of the drugs will be ineffective or too many side effects and development failure, only 10% of the drugs can successfully pass the Food and Drug Administration inspection, was approved for marketing.

As a result, pharmaceutical companies raise the price of successful drugs to compensate for the loss of failed drugs, so the price of some drugs is currently high.

In order to process molecules with such large data and speed up the drug development process, Hannes Stärk, a first-year student in MIT’s Department of Electrical Engineering and Computer Science, has developed a geometric deep learning model called “EquBind. EquBind runs 1200 times faster than the fastest existing molecular computational docking models and is able to find drug-like molecules much faster.

EquBind models can accurately predict protein structures and improve drug discovery efficiency
Most of the traditional molecular docking models find drug-like molecules by a method called “ligand-to-protein binding”. Specifically, the model takes in a large number of sample molecules and then allows the ligand to bind to various molecules, which are then scored by the model and ranked to find the most suitable molecule. However, this is a cumbersome process and the model is less efficient in finding drug-like molecules.

Hannes Stärk gives an analogy to this process, saying, “The typical ‘ligand-protein’ approach used to be like trying to get a model to insert a key into a lock with many keyholes, and the model spends a lot of time scoring the suitability of the key and each of the holes and then selecting the most suitable one. and then picking the one that fits best.”

He goes on to explain, “EquBind skips the most time-consuming step and can predict the most suitable ‘keyhole’ in advance when a new molecule is encountered, which is called ‘blind docking. EquBind has a built-in geometric inference algorithm that helps the model learn the basic structure of the molecule. This algorithm allows EquBind to directly predict the best fit when encountering a new molecule, without spending a lot of time trying different positions and scoring them.”

The model caught the attention of Pat Walters, chief data officer of the therapeutic company Relay. Walters suggested that Hannes Stärk’s research group use the model for drug development used to treat lung cancer, leukemia and gastrointestinal tumors. Typically, protein ligands for drugs in these areas are difficult to dock with most traditional methods, but EquBind allows them to dock successfully.

EquBind offers a unique solution to the protein docking problem by addressing issues such as structure prediction and binding site identification,” said Walters. This approach makes good use of thousands of publicly available crystal structure information, and EquBind could impact the field in new ways.”

Hannes Stärk, author of the paper publishing the technique, which will be accepted at the International Conference on Machine Learning (ICML) in July, said, “I’m looking forward to receiving some ideas for improvements to the EquBind model at this conference.”

The pharmaceutical field is a natural AI scenario. The long cycle, high cost, and low success rate of new drug development leave a huge place for AI: machines can learn data, mine data, summarize and summarize the laws of drug development outside of expert experience, and then optimize all aspects of the drug development process, which not only can improve the efficiency and success rate of drug development, but also is expected to reduce R&D costs and trial and error costs.

Because of such characteristics and development potential, AI drug making is currently gaining momentum. However, there are some industry insiders who say that AI is only playing a supporting role in the pharmaceutical process, and cannot bypass the inherent processes and mechanisms of the industry, so it is impossible to do what we have been doing for 10 years in two or three years.

But overall, there are still new technological breakthroughs in the field of AI pharmaceuticals, and development is booming.

The post 1200x more efficient, MIT develops new model for AI drug making appeared first on TechGoing.