Fullmoonfever
Top 20
Great to see so much happening.
Just wait till formal Ann's start rolling out
Bit going on personally so skimming posts here and there.
Liked the find on ANT61.
Not sure if this posted already but found this article about them on Medium.
Awesome to see the AWS / Amazon accelerator connection as well as Mass Robotics.
While no mention of Akida at this point, you'd like to think ANT61's partners well aware of us through ANT now as well.
AWS accelerating ANT61 development in space
Naturally, as a young company, weāve started by simultaneously running into multiple directions: creating prototypes of our future technology, researching problems worth solving, sizing opportunities and building partnerships.
We have applied to several acceleration programs, and several of them accepted us, so we were fortunate to pick the ones we wanted to do. AWS Robotics, led together by AWS and Mass Robotics from Boston.
Here we would like to share our experience with AWS Robotics and talk about the program and how it challenged and propelled us further than we thought.
Following the program, weāve been selected by Amazon to exhibit our first on-orbit service prototype at Re:MARS 2022 . If you also plan to be there, weād be happy to catch up and talk the AWS Robotics program, our takeaways and future plans. Find us in the Tech Showcase area!
First and foremost, weāve met a great mentor, Matthew Hanson, who has decades of experience in robotic system architecture and is one of the key robotic experts in AWS. His contribution to our success is invaluable and deserves a separate article!
Secondly, weāve built our first digital prototype of the satellite repair robot and have validated our autonomous control system technology in accurate simulation.
The third and precious outcome: weāve tapped into AWS ML engineersā wisdom, significantly reduced our robotās training costs, and took our AI game to the next level.
All of that in the record-break time and with the intense focus!
We have been very fortunate to have the one and only Matthew Hanson.
Matthew Hanson
Matt has a great experience in robotics architecture, everything from control systems, planning, behaviour, autonomy, simulations and of course, everything at AWS that has Robotics.
He understands the startupās unique challenges and has helped us find the quickest path to our goal. Matthew is also very supportive of our cause of sustainable development in space, which means a lot.
As a new business in a conservative industry like space, forming a support group around your team is very important.
Matt was one of the first people who believed in us from the start when we didnāt have much to show except for our prior experience, ambition, hard work and tenacity. He has contributed to everything we did during the program and was kind enough to continue working with us after the program, doing everything he could to help us succeed. We will always remain grateful to Matt, and hopefully, one day, weāll find a great way to reward him in return.
Since the AWS acceleration program is just four weeks, we wonāt have time to develop all systems from scratch. Our main goal was to learn as much as possible about the spacecraftās control systems and how robotic manipulators can be used for docking and repairing the satellite. And weāve decided to learn from the best and used NASAās Astrobee robot that has already flown in ISS as the base for our on-orbit servicing machine.
Astrobee robot on ISS
As a result, weāve been able to apply our ML-based control systems to demonstrate that our robot can autonomously match orbit with a satellite and safely reach out to it.
You can see the simulation video from our AWS Robotics Accelerator Demo day.
At ANT61, we use Deep Reinforcement Learning to teach our robots skills that they can apply for autonomous installation. This approach means that our robots get better at their tasks simply by trying many-many times, rather than us writing slightly better code for them every day.
Reinforcement Learning is still a very new area of Deep Learning. It is akin to gold mining: you have the general directions where to look, but the rich vein can only be found by trying to dig in many places and analysing the results. Our AI engineers always experiment with various algorithms, reward regimes and parameters to get better and more consistent results. The more experiments we can run, the higher our chances of success. On average, one robot takes about two weeks of 24x7 work to learn an essential skill like drilling a hole in a wall with reasonably quality. It will then take several times more training hours to perfect this skill.
Time is our most precious resource, and compressing the time it takes us to determine whether the experiment is successful is very important in our business.
One of the great perks of the accelerator is direct access to senior engineers at AWS. Our mentor, Matthew Hanson, has linked us to the engineers in the AWS Sagemaker team, who provided us with some paths we could try to reduce our costs and run the distributed training of thousands of robots on the same budget.
Now the experiments take only hours instead of weeks, and we can run multiple experiments in parallel without breaking the bank.
Quantity ā© Speed of Innovation
The ability to run multiple experiments, generating years-worth of the data in just a few hours, allowed us to significantly step up our game and use the RL techniques that are usually unavailable to robotics due to the high cost of each data point. Now we can run unsupervised hyper-parameter tuning and neural architecture search.
Essentially, instead of humans coming up with a new experiment to improve our training regimes, AI generates these experiments.
AI that is training other AIs.
As KƔroly Szolnai-FehƩr says, What a time to be alive!
Just wait till formal Ann's start rolling out
Bit going on personally so skimming posts here and there.
Liked the find on ANT61.
Not sure if this posted already but found this article about them on Medium.
Awesome to see the AWS / Amazon accelerator connection as well as Mass Robotics.
While no mention of Akida at this point, you'd like to think ANT61's partners well aware of us through ANT now as well.
ANT61 accelerated by AWS
AWS accelerating ANT61 development in space
ANT61
ANT61 is a robotics & AI startup founded in the second half of 2021. Our main goal is to enable sustainable development in space and back on Earth; we believe that autonomous robots are the way to go about it.Naturally, as a young company, weāve started by simultaneously running into multiple directions: creating prototypes of our future technology, researching problems worth solving, sizing opportunities and building partnerships.
We have applied to several acceleration programs, and several of them accepted us, so we were fortunate to pick the ones we wanted to do. AWS Robotics, led together by AWS and Mass Robotics from Boston.
Here we would like to share our experience with AWS Robotics and talk about the program and how it challenged and propelled us further than we thought.
Following the program, weāve been selected by Amazon to exhibit our first on-orbit service prototype at Re:MARS 2022 . If you also plan to be there, weād be happy to catch up and talk the AWS Robotics program, our takeaways and future plans. Find us in the Tech Showcase area!
The program
AWS Robotics is a 4-week program that balances the workshops on business, technology, marketing and customer acquisition with dedicated 1:1 mentorship sessions. One of the programās goals is to build something new and move your MVP or POC one step forward on technology and product lines.The outcomes
We have received three primary outcomes from this program.First and foremost, weāve met a great mentor, Matthew Hanson, who has decades of experience in robotic system architecture and is one of the key robotic experts in AWS. His contribution to our success is invaluable and deserves a separate article!
Secondly, weāve built our first digital prototype of the satellite repair robot and have validated our autonomous control system technology in accurate simulation.
The third and precious outcome: weāve tapped into AWS ML engineersā wisdom, significantly reduced our robotās training costs, and took our AI game to the next level.
All of that in the record-break time and with the intense focus!
Outcome #1: The mentor
Every startup in the AWS Robotics Accelerator program is assigned a mentor.We have been very fortunate to have the one and only Matthew Hanson.
Matthew Hanson
Matt has a great experience in robotics architecture, everything from control systems, planning, behaviour, autonomy, simulations and of course, everything at AWS that has Robotics.
He understands the startupās unique challenges and has helped us find the quickest path to our goal. Matthew is also very supportive of our cause of sustainable development in space, which means a lot.
As a new business in a conservative industry like space, forming a support group around your team is very important.
Matt was one of the first people who believed in us from the start when we didnāt have much to show except for our prior experience, ambition, hard work and tenacity. He has contributed to everything we did during the program and was kind enough to continue working with us after the program, doing everything he could to help us succeed. We will always remain grateful to Matt, and hopefully, one day, weāll find a great way to reward him in return.
Outcome #2: The space MVP
The period of our AWS Robotics acceleration program coincided with our Space Accelerator run with the support of the Australian Space Agency in Adelaide. We were actively looking into the on-orbit repair and servicing and thought it would be great to build the first virtual prototype of our future space robot with AWS.Since the AWS acceleration program is just four weeks, we wonāt have time to develop all systems from scratch. Our main goal was to learn as much as possible about the spacecraftās control systems and how robotic manipulators can be used for docking and repairing the satellite. And weāve decided to learn from the best and used NASAās Astrobee robot that has already flown in ISS as the base for our on-orbit servicing machine.
Astrobee robot on ISS
As a result, weāve been able to apply our ML-based control systems to demonstrate that our robot can autonomously match orbit with a satellite and safely reach out to it.
You can see the simulation video from our AWS Robotics Accelerator Demo day.
Outcome #3: The next-level AI training
The price of the experimentAt ANT61, we use Deep Reinforcement Learning to teach our robots skills that they can apply for autonomous installation. This approach means that our robots get better at their tasks simply by trying many-many times, rather than us writing slightly better code for them every day.
Reinforcement Learning is still a very new area of Deep Learning. It is akin to gold mining: you have the general directions where to look, but the rich vein can only be found by trying to dig in many places and analysing the results. Our AI engineers always experiment with various algorithms, reward regimes and parameters to get better and more consistent results. The more experiments we can run, the higher our chances of success. On average, one robot takes about two weeks of 24x7 work to learn an essential skill like drilling a hole in a wall with reasonably quality. It will then take several times more training hours to perfect this skill.
Time is our most precious resource, and compressing the time it takes us to determine whether the experiment is successful is very important in our business.
AWS=scale
Even before the accelerator, we have already been using multi-agent training, where experience from several robots is accumulated to train one brain, which is then deployed to the whole fleet, and the training continues. However, our budget severely limited how many robots we could run in parallel.One of the great perks of the accelerator is direct access to senior engineers at AWS. Our mentor, Matthew Hanson, has linked us to the engineers in the AWS Sagemaker team, who provided us with some paths we could try to reduce our costs and run the distributed training of thousands of robots on the same budget.
Now the experiments take only hours instead of weeks, and we can run multiple experiments in parallel without breaking the bank.
Quantity ā© Speed of Innovation
The ability to run multiple experiments, generating years-worth of the data in just a few hours, allowed us to significantly step up our game and use the RL techniques that are usually unavailable to robotics due to the high cost of each data point. Now we can run unsupervised hyper-parameter tuning and neural architecture search.
Essentially, instead of humans coming up with a new experiment to improve our training regimes, AI generates these experiments.
AI that is training other AIs.
As KƔroly Szolnai-FehƩr says, What a time to be alive!