This is the perfect opportunity to pitch my own version of this robot! Implemented in a ROS framework, with a working walking gait, and some decent documentation: https://github.com/mike4192/spotMicro
I stumbled on nicrusso7's github page as I was doing my own initial research. I initially, naively, thought I could easily expand on his work and reinforcement learn my way to a working walking gait pretty quickly. However once I got a feel for the hardware and it's performance limits and the play in the system, I realized differences between the simulated model and real life would be significant, and would make debugging the translation from simulation to real life difficult. Also I'm teaching myself software development and reinforcement learning was another whole thing to learn.
So instead I took a different approach and implemented a more "conventional" gait, with inspiration and examples from other similar projects.
I've abandoned reinforcement learning for now, as I'm more interested in implementing mapping and motion planning going forward. But one day I'd like to revisit it and see how the pipeline for ML to real implementation works for something like this low level control.
Well done!
Btw I do agree with you regarding the knowledge transfer on the real robot platform. The open source design is built to work with very cheap servos and the alinement is not quite easy too.
As you, I was totally new to RL when I started (almost 1 year ago) - I’m a DevOps engineer during the day :)
The goal of this project for me is learning ML more than build a product - that’s why the limited focus on hw implementation. Anyway, there are examples on the web on how to run the policies on the real world robot (even using ROS) - maybe in the future I’ll digging in this topic as well!
Reading the comments it struck me: why is tele operated agricultural robots not a thing? A robot with wheels and an array of cameras and arms. It rolls down the line and remote workers pick the fruit. Maybe even out of country workers being paid pennies...I’m sure the latency could be overcome. You could host thousands of humans with very few robots. Better productivity, lower cost in the long run and you’re resilient against labor problems which is really good when all your labor is illegal. If nobody gives a good reason why it doesn’t exist I would be receptive to people who might want to do a start up...
Googling around I find some recent stuff. It seems to have started being on people’s minds in the past five years only. A lot of complicated articulated arms and inputs. I am only more convinced that there is arbitrage here... a very simple articulation mechanism, something akin to what hello robot did vs willow garage, would be totally new. And you don’t have to replicate perfectly the performance of humans because it’s a sliding scale, you could use the robot to downsize at first, following it with a small human team.
Depends on what you’re harvesting but the challenges are likely to be in the actual picking motion not dropping/bruising/cutting the food, if there is any selection happening the senses of touch and smell are largely gone, navigating the plant without damaging it, not compacting the soil around the roots more than necessary, overall speed, etc.
Bespoke harvesters for things like potatoes and almonds obviously destroy humans in the productivity department. If we’re still picking it that’s a good signal that automation will be tricky.
We probably have the technology to do most of it now, but you may wind up with a half million dollar robot.
Machines get stuck in mud, trample / cut plants, pieces of plants get stuck in machine, and require frequent maintenance. The more degrees of freedom, the more failure points. Also, robots are $$$$, while unskilled human labor is $. Still, there several agro robotics startups.
Few years ago I was part of the networking team to build a new greenhouse. Normally they would need hundreds of employees but this was reduced to 3. Fully automated, everything on rails.
This is how it is already. It just takes time to amortize, build, and operate facilities.
I have been fantasizing about a similar setup for some time, but an arachnid instead of a quadrupod. I imagined it to pick up trash in land with a lot of vegetation. Never got around it because I am on my own personal project and this sounds like a massive amount of work for a single person. Congrats and good luck. Will bookmark it so I can get back to it when I am freer. Sounds like a very good starting point for this topic.
One thing, why use a machine learning algorithm for path finding in the maze instead of just using an actual path finding algo? Did I miss something?
I've thought about a similar platform, but as an agricultural robot for smallholding.
If you have a quarter acre, you can have a nice garden, and it's kind of labor intensive but hopefully enjoyable, and you can get some nice food out of it.
With five to ten acres, food security is in reach, and some left over for market. But it's really rather a lot of work, and the tools farmers use to make that work easy are massively oversized for the job, and require you to lay out the land according to the logic of a much larger farm.
A pig-sized arachnid robot, with a vision module and a few replaceable tools for 'mandibles', a replaceable lithium pack for an abdomen: that could bring the amount of work back down to a part-time job.
This is a company, not a project. But backed with open-source software so that the machine learning and supplemental tools can be improved by users, I think it could be quite a successful one, and make the world more robust and anti-fragile.
Arachnid robots would be ideal for household vacuum cleaners. Roomba's are useless if you have cables on the floor, lots of equipment and furniture, stairs that need vacuuming, roughly surfaced floors, or a variety of other conditions.
Re the maze, I've got just the terrain so far - I'm working on the gym environment.. I was thinking to write 2 different versions of it, one in which the robot has a map of the maze and navigate it using a path finding algo (probably A*) and another version without any map (I'll probably need a lidar for this).
This looks very cool! I wonder how close it is to the method that Boston Dynamics uses to control the their version. Did they simply train a machine learning algorithm for way longer than a hobbyist can or is there a ton of manual tweaking?
Boston Dynamics (famously) does not use machine learning to control their robots. Maybe a little bit of computer vision, but all control is with more classical approaches.
Boston Dynamics probably does some form of force control. That is controlling the force applied to the ground by the legs in addition to just the position the leg is moving. The actuators here appear to be purely position controlled.
thanks! well, good question. I think it's a mix between ML (probably not just simulations but also directly learn from real-world experiences) and more classic approaches. There are some very interesting examples on the Google AI blog like this one (https://ai.googleblog.com/2020/04/exploring-nature-inspired-...)
mike4921|5 years ago
I stumbled on nicrusso7's github page as I was doing my own initial research. I initially, naively, thought I could easily expand on his work and reinforcement learn my way to a working walking gait pretty quickly. However once I got a feel for the hardware and it's performance limits and the play in the system, I realized differences between the simulated model and real life would be significant, and would make debugging the translation from simulation to real life difficult. Also I'm teaching myself software development and reinforcement learning was another whole thing to learn.
So instead I took a different approach and implemented a more "conventional" gait, with inspiration and examples from other similar projects.
I've abandoned reinforcement learning for now, as I'm more interested in implementing mapping and motion planning going forward. But one day I'd like to revisit it and see how the pipeline for ML to real implementation works for something like this low level control.
nicrusso7|5 years ago
As you, I was totally new to RL when I started (almost 1 year ago) - I’m a DevOps engineer during the day :) The goal of this project for me is learning ML more than build a product - that’s why the limited focus on hw implementation. Anyway, there are examples on the web on how to run the policies on the real world robot (even using ROS) - maybe in the future I’ll digging in this topic as well!
vwat|5 years ago
Googling around I find some recent stuff. It seems to have started being on people’s minds in the past five years only. A lot of complicated articulated arms and inputs. I am only more convinced that there is arbitrage here... a very simple articulation mechanism, something akin to what hello robot did vs willow garage, would be totally new. And you don’t have to replicate perfectly the performance of humans because it’s a sliding scale, you could use the robot to downsize at first, following it with a small human team.
jcims|5 years ago
Bespoke harvesters for things like potatoes and almonds obviously destroy humans in the productivity department. If we’re still picking it that’s a good signal that automation will be tricky.
We probably have the technology to do most of it now, but you may wind up with a half million dollar robot.
ZitchDog|5 years ago
tintor|5 years ago
sleepysysadmin|5 years ago
This is how it is already. It just takes time to amortize, build, and operate facilities.
nicrusso7|5 years ago
ptsneves|5 years ago
One thing, why use a machine learning algorithm for path finding in the maze instead of just using an actual path finding algo? Did I miss something?
samatman|5 years ago
If you have a quarter acre, you can have a nice garden, and it's kind of labor intensive but hopefully enjoyable, and you can get some nice food out of it.
With five to ten acres, food security is in reach, and some left over for market. But it's really rather a lot of work, and the tools farmers use to make that work easy are massively oversized for the job, and require you to lay out the land according to the logic of a much larger farm.
A pig-sized arachnid robot, with a vision module and a few replaceable tools for 'mandibles', a replaceable lithium pack for an abdomen: that could bring the amount of work back down to a part-time job.
This is a company, not a project. But backed with open-source software so that the machine learning and supplemental tools can be improved by users, I think it could be quite a successful one, and make the world more robust and anti-fragile.
rapjr9|5 years ago
nicrusso7|5 years ago
Re the maze, I've got just the terrain so far - I'm working on the gym environment.. I was thinking to write 2 different versions of it, one in which the robot has a map of the maze and navigate it using a path finding algo (probably A*) and another version without any map (I'll probably need a lidar for this).
hkt|5 years ago
(Obviously this isn't very serious. I do however think this project is great)
nicrusso7|5 years ago
xVedun|5 years ago
317070|5 years ago
gene-h|5 years ago
nicrusso7|5 years ago