Tag Archives: perfect

#439089 Ingenuity’s Chief Pilot Explains How ...

Posted on April 9, 2021 by Android

On April 11, the Mars helicopter Ingenuity will take to the skies of Mars for the first time. It will do so fully autonomously, out of necessity—the time delay between Ingenuity’s pilots at the Jet Propulsion Laboratory and Jezero Crater on Mars makes manual or even supervisory control impossible. So the best that the folks at JPL can do is practice as much as they can in simulation, and then hope that the helicopter can handle everything on its own.

Here on Earth, simulation is a critical tool for many robotics applications, because it doesn’t rely on access to expensive hardware, is non-destructive, and can be run in parallel and at faster-than-real-time speeds to focus on solving specific problems. Once you think you’ve gotten everything figured out in simulation, you can always give it a try on the real robot and see how close you came. If it works in real life, great! And if not, well, you can tweak some stuff in the simulation and try again.

For the Mars helicopter, simulation is much more important, and much higher stakes. Testing the Mars helicopter under conditions matching what it’ll find on Mars is not physically possible on Earth. JPL has flown engineering models in Martian atmospheric conditions, and they’ve used an actuated tether to mimic Mars gravity, but there’s just no way to know what it’ll be like flying on Mars until they’ve actually flown on Mars. With that in mind, the Ingenuity team has been relying heavily on simulation, since that’s one of the best tools they have to prepare for their Martian flights. We talk with Ingenuity’s Chief Pilot, Håvard Grip, to learn how it all works.

Ingenuity Facts:
Body Size: a box of tissues

Brains: Qualcomm Snapdragon 801

Weight: 1.8 kilograms

Propulsion: Two 1.2m carbon fiber rotors

Navigation sensors: VGA camera, laser altimeter, inclinometer

Ingenuity is scheduled to make its first flight no earlier than April 11. Before liftoff, the Ingenuity team will conduct a variety of pre-flight checks, including verifying the responsiveness of the control system and spinning the blades up to full speed (2,537 rpm) without lifting off. If everything looks good, the first flight will consist of a 1 meter per second climb to 3 meters, 30 seconds of hover at 3 meters while rotating in place a bit, and then a descent to landing. If Ingenuity pulls this off, that will have made its entire mission a success. There will be more flights over the next few weeks, but all it takes is one to prove that autonomous helicopter flight on Mars is possible.

Last month, we spoke with Mars Helicopter Operations Lead Tim Canham about Ingenuity’s hardware, software, and autonomy, but we wanted to know more about how the Ingenuity team has been using simulation for everything from vehicle design to flight planning. To answer our questions, we talked with JPL’s Håvard Grip, who led the development of Ingenuity’s navigation and flight control systems. Grip also has the title of Ingenuity Chief Pilot, which is pretty awesome. He summarizes this role as “operating the flight control system to make the helicopter do what we want it to do.”

IEEE Spectrum: Can you tell me about the simulation environment that JPL uses for Ingenuity’s flight planning?

Håvard Grip: We developed a Mars helicopter simulation ourselves at JPL, based on a multi-body simulation framework that’s also developed at JPL, called DARTS/DSHELL. That's a system that has been in development at JPL for about 30 years now, and it's been used in a number of missions. And so we took that multibody simulation framework, and based on it we built our own Mars helicopter simulation, put together our own rotor model, our own aerodynamics models, and everything else that's needed in order to simulate a helicopter. We also had a lot of help from the rotorcraft experts at NASA Ames and NASA Langley.

Image: NASA/JPL

Ingenuity in JPL’s flight simulator.

Without being able to test on Mars, how much validation are you able to do of what you’re seeing in simulation?

We can do a fair amount, but it requires a lot of planning. When we made our first real prototype (with a full-size rotor that looked like what we were thinking of putting on Mars) we first spent a lot of time designing it and using simulation tools to guide that design, and when we were sufficiently confident that we were close enough, and that we understood enough about it, then we actually built the thing and designed a whole suite of tests in a vacuum chamber where where we could replicate Mars atmospheric conditions. And those tests were before we tried to fly the helicopter—they were specifically targeted at what we call system identification, which has to do with figuring out what the true properties, the true dynamics of a system are, compared to what we assumed in our models. So then we got to see how well our models did, and in the places where they needed adjustment, we could go back and do that.

The simulation work that we really started after that very first initial lift test, that’s what allowed us to unlock all of the secrets to building a helicopter that can fly on Mars.
—Håvard Grip, Ingenuity Chief Pilot

We did a lot of this kind of testing. It was a big campaign, in several stages. But there are of course things that you can't fully replicate, and you do depend on simulation to tie things together. For example, we can't truly replicate Martian gravity on Earth. We can replicate the atmosphere, but not the gravity, and so we have to do various things when we fly—either make the helicopter very light, or we have to help it a little bit by pulling up on it with a string to offload some of the weight. These things don't fully replicate what it will be like on Mars. We also can't simultaneously replicate the Mars aerodynamic environment and the physical and visual surroundings that the helicopter will be flying in. These are places where simulation tools definitely come in handy, with the ability to do full flight tests from A to B, with the helicopter taking off from the ground, running the flight software that it will be running on board, simulating the images that the navigation camera takes of the ground below as it flies, feeding that back into the flight software, and then controlling it.

To what extent can simulation really compensate for the kinds of physical testing that you can’t do on Earth?

It gives you a few different possibilities. We can take certain tests on Earth where we replicate key elements of the environment, like the atmosphere or the visual surroundings for example, and you can validate your simulation on those parameters that you can test on Earth. Then, you can combine those things in simulation, which gives you the ability to set up arbitrary scenarios and do lots and lots of tests. We can Monte Carlo things, we can do a flight a thousand times in a row, with small perturbations of various parameters and tease out what our sensitivities are to those things. And those are the kinds of things that you can't do with physical tests, both because you can't fully replicate the environment and also because of the resources that would be required to do the same thing a thousand times in a row.

Because there are limits to the physical testing we can do on Earth, there are elements where we know there's more uncertainty. On those aspects where the uncertainty is high, we tried to build in enough margin that we can handle a range of things. And simulation gives you the ability to then maybe play with those parameters, and put them at their outer limits, and test them beyond where the real parameters are going to be to make sure that you have robustness even in those extreme cases.

How do you make sure you’re not relying on simulation too much, especially since in some ways it’s your only option?

It’s about anchoring it in real data, and we’ve done a lot of that with our physical testing. I think what you’re referring to is making your simulation too perfect, and we’re careful to model the things that matter. For example, the simulated sensors that we use have realistic levels of simulated noise and bias in them, the navigation camera images have realistic levels of degradation, we have realistic disturbances from wind gusts. If you don’t properly account for those things, then you’re missing important details. So, we try to be as accurate as we can, and to capture that by overbounding in areas where we have a high degree of uncertainty.

What kinds of simulated challenges have you put the Mars helicopter through, and how do you decide how far to push those challenges?

One example is that we can simulate going over rougher terrain. We can push that, and see how far we can go and still have the helicopter behave the way that we want it to. Or we can inject levels of noise that maybe the real sensors don't see, but you want to just see how far you can push things and make sure that it's still robust.

Where we put the limits on this and what we consider to be realistic is often a challenge. We consider this on a case by case basis—if you have a sensor that you're dealing with, you try to do testing with it to characterize it and understand its performance as much as possible, and you build a level of confidence in it that allows you to find the proper balance.

When it comes to things like terrain roughness, it's a little bit of a different thing, because we're actually picking where we're flying the helicopter. We have made that choice, and we know what the terrain looks like around us, so we don’t have to wonder about that anymore.

Image: NASA/JPL-Caltech/University of Arizona

Satellite image of the Ingenuity flight area.

The way that we’re trying to approach this operationally is that we should be done with the engineering at this point. We’re not depending on going back and resimulating things, other than a few checks here and there.

Are there any examples of things you learned as part of the simulation process that resulted in changes to the hardware or mission?

You know, it’s been a journey. One of the early things that we discovered as part of modeling the helicopter was that the rotor dynamics were quite different for a helicopter on Mars, in particular with respect to how the rotor responds to the up and down bending of the blades because they’re not perfectly rigid. That motion is a very important influence on the overall flight dynamics of the helicopter, and what we discovered as we started modeling was that this motion is damped much less on Mars. Under-damped oscillatory things like that, you kind of figure might pose a control issue, and that is the case here: if you just naively design it as you might a helicopter on Earth, without taking this into account, you could have a system where the response to control inputs becomes very sluggish. So that required changes to the vehicle design from some of the very early concepts, and it led us to make a rotor that’s extremely light and rigid.

The design cycle for the Mars helicopter—it’s not like we could just build something and take it out to the back yard and try it and then come back and tweak it if it doesn’t work. It’s a much bigger effort to build something and develop a test program where you have to use a vacuum chamber to test it. So you really want to get as close as possible up front, on your first iteration, and not have to go back to the drawing board on the basic things.

So how close were you able to get on your first iteration of the helicopter design?

[This video shows] a very early demo which was done more or less just assuming that things were going to behave as they would on Earth, and that we’d be able to fly in a Martian atmosphere just spinning the rotor faster and having a very light helicopter. We were basically just trying to demonstrate that we could produce enough lift. You can see the helicopter hopping around, with someone trying to joystick it, but it turned out to be very hard to control. This was prior to doing any of the modeling that I talked about earlier. But once we started seriously focusing on the modeling and simulation, we then went on to build a prototype vehicle which had a full-size rotor that’s very close to the rotor that will be flying on Mars. One difference is that prototype had cyclic control only on the lower rotor, and later we added cyclic control on the upper rotor as well, and that decision was informed in large part by the work we did in simulation—we’d put in the kinds of disturbances that we thought we might see on Mars, and decided that we needed to have the extra control authority.

How much room do you think there is for improvement in simulation, and how could that help you in the future?

The tools that we have were definitely sufficient for doing the job that we needed to do in terms of building a helicopter that can fly on Mars. But simulation is a compute-intensive thing, and so I think there’s definitely room for higher fidelity simulation if you have the compute power to do so. For a future Mars helicopter, you could get some benefits by more closely coupling together high-fidelity aerodynamic models with larger multi-body models, and doing that in a fast way, where you can iterate quickly. There’s certainly more potential for optimizing things.

Photo: NASA/JPL-Caltech

Ingenuity preparing for flight.

Watching Ingenuity’s first flight take place will likely be much like watching the Perseverance landing—we’ll be able to follow along with the Ingenuity team while they send commands to the helicopter and receive data back, although the time delay will mean that any kind of direct control won’t be possible. If everything goes the way it’s supposed to, there will hopefully be some preliminary telemetry from Ingenuity saying so, but it sounds like we’ll likely have to wait until April 12 before we get pictures or video of the flight itself.

Because Mars doesn’t care what time it is on Earth, the flight will actually be taking place very early on April 12, with the JPL Mission Control livestream starting at 3:30 a.m. EDT (12:30 a.m. PDT). Details are here. Continue reading →

Posted in Human Robots

#439032 To Learn To Deal With Uncertainty, This ...

Posted on March 26, 2021 by Android

AI is endowing robots, autonomous vehicles and countless of other forms of tech with new abilities and levels of self-sufficiency. Yet these models faithfully “make decisions” based on whatever data is fed into them, which could have dangerous consequences. For instance, if an autonomous car is driving down a highway and the sensor picks up a confusing signal (e.g., a paint smudge that is incorrectly interpreted as a lane marking), this could cause the car to swerve into another lane unnecessarily.

But in the ever-evolving world of AI, researchers are developing new ways to address challenges like this. One group of researchers has devised a new algorithm that allows the AI model to account for uncertain data, which they describe in a study published February 15 in IEEE Transactions on Neural Networks and Learning Systems.

“While we would like robots to work seamlessly in the real world, the real world is full of uncertainty,” says Michael Everett, a post-doctoral associate at MIT who helped develop the new approach. “It's important for a system to be aware of what it knows and what it is unsure about, which has been a major challenge for modern AI.”

His team focused on a type of AI called reinforcement learning (RL), whereby the model tries to learn the “value” of taking each action in a given scenario through trial-and-error. They developed a secondary algorithm, called Certified Adversarial Robustness for deep RL (CARRL), that can be built on top of an existing RL model.

“Our key innovation is that rather than blindly trusting the measurements, as is done today [by AI models], our algorithm CARRL thinks through all possible measurements that could have been made, and makes a decision that considers the worst-case outcome,” explains Everett.

In their study, the researchers tested CARRL across several different tasks, including collision avoidance simulations and Atari pong. For younger readers who may not be familiar with it, Atari pong is a classic computer game whereby an electronic paddle is used to direct a ping pong on the screen. In the test scenario, CARRL helped move the paddle slightly higher or lower to compensate for the possibility that the ball could approach at a slightly different point than what the input data indicated. All the while, CARRL would try to ensure that the ball would make contact with at least some part of paddle.

Gif: MIT Aerospace Controls Laboratory

In a perfect world, the information that an AI model is fed would be accurate all the time and AI model will perform well (left). But in some cases, the AI may be given inaccurate data, causing it to miss its targets (middle). The new algorithm CARRL helps AIs account for uncertainty in its data inputs, yielding a better performance when relying on poor data (right).

Across all test scenarios, the RL model was better at compensating for potential inaccurate or “noisy” data with CARRL, than without CARRL.

But the results also show that, like with humans, too much self-doubt and uncertainty can be unhelpful. In the collision avoidance scenario, for example, indulging in too much uncertainty caused the main moving object in the simulation to avoid both the obstacle and its goal. “There is definitely a limit to how ‘skeptical’ the algorithm can be without becoming overly conservative,” Everett says.

This research was funded by Ford Motor Company, but Everett notes that it could be applicable under many other commercial applications requiring safety-aware AI, including aerospace, healthcare, or manufacturing domains.

“This work is a step toward my vision of creating ‘certifiable learning machines’—systems that can discover how to explore and perform in the real world on their own, while still having safety and robustness guarantees,” says Everett. “We'd like to bring CARRL into robotic hardware while continuing to explore the theoretical challenges at the interface of robotics and AI.” Continue reading →

Posted in Human Robots

#439012 Video Friday: Man-Machine Synergy ...

Posted on March 20, 2021 by Android

Video Friday is your weekly selection of awesome robotics videos, collected by your Automaton bloggers. We’ll also be posting a weekly calendar of upcoming robotics events for the next few months; here's what we have so far (send us your events!):

RoboSoft 2021 – April 12-16, 2021 – [Online Conference]
ICRA 2021 – May 30-5, 2021 – Xi'an, China
DARPA SubT Finals – September 21-23, 2021 – Louisville, KY, USA
WeRobot 2021 – September 23-25, 2021 – Coral Gables, FL, USA
Let us know if you have suggestions for next week, and enjoy today's videos.

Man-Machine Synergy Effectors, Inc. is a Japanese company working on an absolutely massive “human machine synergistic effect device,” which is a huge robot controlled by a nearby human using a haptic rig.

From the look of things, the next generation will be able to move around. Whoa.

[ MMSE ]

This method of loading and unloading AMRs without having them ever stop moving is so obvious that there must be some equally obvious reason why I've never seen it done in practice.

The LoadRunner is able to transport and sort parcels weighing up to 30 kilograms. This makes it the perfect luggage carrier for airports. These AI-driven go-carts can also work in concert as larger collectives to carry large, heavy and bulky objects. Every LoadRunner can also haul up to four passive trailers. Powered by four electric motors, the LoadRunner sharply brakes at just the right moment right in front of its destination and the payload slides from the robot onto the delivery platform.

[ Fraunhofer ] via [ Gizmodo ]

Ayato Kanada at Kyushu University wrote in to share this clever “dislocatable joint,” a way of combining continuum and rigid robots.

[ Paper ]

Thanks Ayato!

The DodgeDrone challenge revisits the popular dodgeball game in the context of autonomous drones. Specifically, participants will have to code navigation policies to fly drones between waypoints while avoiding dynamic obstacles. Drones are fast but fragile systems: as soon as something hits them, they will crash! Since objects will move towards the drone with different speeds and acceleration, smart algorithms are required to avoid them!

This could totally happen in real life, and we need to be prepared for it!

[ DodgeDrone Challenge ]

In addition to winning the Best Student Design Competition CREATIVITY Award at HRI 2021, this paper would also have won the Best Paper Title award, if that award existed.

[ Paper ]

Robots are traditionally bound by a fixed morphology during their operational lifetime, which is limited to adapting only their control strategies. Here we present the first quadrupedal robot that can morphologically adapt to different environmental conditions in outdoor, unstructured environments.

We show that the robot exploits its training to effectively transition between different morphological configurations, exhibiting substantial performance improvements over a non-adaptive approach. The demonstrated benefits of real-world morphological adaptation demonstrate the potential for a new embodied way of incorporating adaptation into future robotic designs.

[ Nature ]

A drone video shot in a Minneapolis bowling alley was hailed as an instant classic. One Hollywood veteran said it “adds to the language and vocabulary of cinema.” One IEEE Spectrum editor said “hey that's pretty cool.”

[ Bryant Lake Bowl ]

It doesn't take a robot to convince me to buy candy, but I think if I buy candy from Relay it's a business expense, right?

[ RIS ]

DARPA is making progress on its AI dogfighting program, with physical flight tests expected this year.

[ DARPA ACE ]

Unitree Robotics has realized that the Empire needs to be overthrown!

[ Unitree ]

Windhover Labs, an emerging leader in open and reliable flight software and hardware, announces the upcoming availability of its first hardware product, a low cost modular flight computer for commercial drones and small satellites.

[ Windhover ]

As robots and autonomous systems are poised to become part of our everyday lives, the University of Michigan and Ford are opening a one-of-a-kind facility where they’ll develop robots and roboticists that help make lives better, keep people safer and build a more equitable society.

[ U Michigan ]

The adaptive robot Rizon combined with a new hybrid electrostatic and gecko-inspired gripping pad developed by Stanford BDML can manipulate bulky, non-smooth items in the most effort-saving way, which broadens the applications in retail and household environments.

[ Flexiv ]

Thanks Yunfan!

I don't know why anyone would want things to get MORE icy, but if you do for some reason, you can make it happen with a Husky.

Is winter over yet?

[ Clearpath ]

Skip ahead to about 1:20 to see a pair of Gita robots following a Spot following a human like a chain of lil’ robot duckings.

[ PFF ]

Here are a couple of retro robotics videos, one showing teleoperated humanoids from 2000, and the other showing a robotic guide dog from 1976 (!)

[ Tachi Lab ]

Thanks Fan!

If you missed Chad Jenkins' talk “That Ain’t Right: AI Mistakes and Black Lives” last time, here's another opportunity to watch from Robotics Today, and it includes a top notch panel discussion at the end.

[ Robotics Today ]

Since its founding in 1979, the Robotics Institute (RI) at Carnegie Mellon University has been leading the world in robotics research and education. In the mid 1990s, RI created NREC as the applied R&D center within the Institute with a specific mission to apply robotics technology in an impactful way on real-world applications. In this talk, I will go over numerous R&D programs that I have led at NREC in the past 25 years.

[ CMU ] Continue reading →

Posted in Human Robots

#437445 Humanoid Robotic Barista

Posted on March 16, 2021 by Android

The “Robotic Coffee Master” by OrionStar is a robot coffee brewing guru that uses expert brewing techniques with millimeter-level stability and accuracy to make the perfect cuppa every time. This robotic barista incorporates 3,000 hours of A.I. learning, 30,000 hours … Continue reading →

Posted in Human Robots

#438801 This AI Thrashes the Hardest Atari Games ...

Posted on March 4, 2021 by Android

Learning from rewards seems like the simplest thing. I make coffee, I sip coffee, I’m happy. My brain registers “brewing coffee” as an action that leads to a reward.

That’s the guiding insight behind deep reinforcement learning, a family of algorithms that famously smashed most of Atari’s gaming catalog and triumphed over humans in strategy games like Go. Here, an AI “agent” explores the game, trying out different actions and registering ones that let it win.

Except it’s not that simple. “Brewing coffee” isn’t one action; it’s a series of actions spanning several minutes, where you’re only rewarded at the very end. By just tasting the final product, how do you learn to fine-tune grind coarseness, water to coffee ratio, brewing temperature, and a gazillion other factors that result in the reward—tasty, perk-me-up coffee?

That’s the problem with “sparse rewards,” which are ironically very abundant in our messy, complex world. We don’t immediately get feedback from our actions—no video-game-style dings or points for just grinding coffee beans—yet somehow we’re able to learn and perform an entire sequence of arm and hand movements while half-asleep.

This week, researchers from UberAI and OpenAI teamed up to bestow this talent on AI.

The trick is to encourage AI agents to “return” to a previous step, one that’s promising for a winning solution. The agent then keeps a record of that state, reloads it, and branches out again to intentionally explore other solutions that may have been left behind on the first go-around. Video gamers are likely familiar with this idea: live, die, reload a saved point, try something else, repeat for a perfect run-through.

The new family of algorithms, appropriately dubbed “Go-Explore,” smashed notoriously difficult Atari games like Montezuma’s Revenge that were previously unsolvable by its AI predecessors, while trouncing human performance along the way.

It’s not just games and digital fun. In a computer simulation of a robotic arm, the team found that installing Go-Explore as its “brain” allowed it to solve a challenging series of actions when given very sparse rewards. Because the overarching idea is so simple, the authors say, it can be adapted and expanded to other real-world problems, such as drug design or language learning.

Growing Pains
How do you reward an algorithm?

Rewards are very hard to craft, the authors say. Take the problem of asking a robot to go to a fridge. A sparse reward will only give the robot “happy points” if it reaches its destination, which is similar to asking a baby, with no concept of space and danger, to crawl through a potential minefield of toys and other obstacles towards a fridge.

“In practice, reinforcement learning works very well, if you have very rich feedback, if you can tell, ‘hey, this move is good, that move is bad, this move is good, that move is bad,’” said study author Joost Huinzinga. However, in situations that offer very little feedback, “rewards can intentionally lead to a dead end. Randomly exploring the space just doesn’t cut it.”

The other extreme is providing denser rewards. In the same robot-to-fridge example, you could frequently reward the bot as it goes along its journey, essentially helping “map out” the exact recipe to success. But that’s troubling as well. Over-holding an AI’s hand could result in an extremely rigid robot that ignores new additions to its path—a pet, for example—leading to dangerous situations. It’s a deceptive AI solution that seems effective in a simple environment, but crashes in the real world.

What we need are AI agents that can tackle both problems, the team said.

Intelligent Exploration
The key is to return to the past.

For AI, motivation usually comes from “exploring new or unusual situations,” said Huizinga. It’s efficient, but comes with significant downsides. For one, the AI agent could prematurely stop going back to promising areas because it thinks it had already found a good solution. For another, it could simply forget a previous decision point because of the mechanics of how it probes the next step in a problem.

For a complex task, the end result is an AI that randomly stumbles around towards a solution while ignoring potentially better ones.

“Detaching from a place that was previously visited after collecting a reward doesn’t work in difficult games, because you might leave out important clues,” Huinzinga explained.

Go-Explore solves these problems with a simple principle: first return, then explore. In essence, the algorithm saves different approaches it previously tried and loads promising save points—once more likely to lead to victory—to explore further.

Digging a bit deeper, the AI stores screen caps from a game. It then analyzes saved points and groups images that look alike as a potential promising “save point” to return to. Rinse and repeat. The AI tries to maximize its final score in the game, and updates its save points when it achieves a new record score. Because Atari doesn’t usually allow people to revisit any random point, the team used an emulator, which is a kind of software that mimics the Atari system but with custom abilities such as saving and reloading at any time.

The trick worked like magic. When pitted against 55 Atari games in the OpenAI gym, now commonly used to benchmark reinforcement learning algorithms, Go-Explore knocked out state-of-the-art AI competitors over 85 percent of the time.

It also crushed games previously unbeatable by AI. Montezuma’s Revenge, for example, requires you to move Pedro, the blocky protagonist, through a labyrinth of underground temples while evading obstacles such as traps and enemies and gathering jewels. One bad jump could derail the path to the next level. It’s a perfect example of sparse rewards: you need a series of good actions to get to the reward—advancing onward.

Go-Explore didn’t just beat all levels of the game, a first for AI. It also scored higher than any previous record for reinforcement learning algorithms at lower levels while toppling the human world record.

Outside a gaming environment, Go-Explore was also able to boost the performance of a simulated robot arm. While it’s easy for humans to follow high-level guidance like “put the cup on this shelf in a cupboard,” robots often need explicit training—from grasping the cup to recognizing a cupboard, moving towards it while avoiding obstacles, and learning motions to not smash the cup when putting it down.

Here, similar to the real world, the digital robot arm was only rewarded when it placed the cup onto the correct shelf, out of four possible shelves. When pitted against another algorithm, Go-Explore quickly figured out the movements needed to place the cup, while its competitor struggled with even reliably picking the cup up.

Combining Forces
By itself, the “first return, then explore” idea behind Go-Explore is already powerful. The team thinks it can do even better.

One idea is to change the mechanics of save points. Rather than reloading saved states through the emulator, it’s possible to train a neural network to do the same, without needing to relaunch a saved state. It’s a potential way to make the AI even smarter, the team said, because it can “learn” to overcome one obstacle once, instead of solving the same problem again and again. The downside? It’s much more computationally intensive.

Another idea is to combine Go-Explore with an alternative form of learning, called “imitation learning.” Here, an AI observes human behavior and mimics it through a series of actions. Combined with Go-Explore, said study author Adrien Ecoffet, this could make more robust robots capable of handling all the complexity and messiness in the real world.

To the team, the implications go far beyond Go-Explore. The concept of “first return, then explore” seems to be especially powerful, suggesting “it may be a fundamental feature of learning in general.” The team said, “Harnessing these insights…may be essential…to create generally intelligent agents.”

Image Credit: Adrien Ecoffet, Joost Huizinga, Joel Lehman, Kenneth O. Stanley, and Jeff Clune Continue reading →

Posted in Human Robots