Roboschool: open-source software for robot simulation | Hacker News Reader

Roboschool: open-source software for robot simulation

blog.openai.com

244 pointsdpf26 comments

Open articleView on HN

Related is NVIDIA's recent announcement of their Isaac Robot Simulator:

https://www.youtube.com/watch?v=oa__wkSmWUw

In addition to that great news, MuJoCo student licenses are now free: http://www.mujoco.org/index.html

Interesting...

One robotics simulator that never seems to get any love is this one:

https://assemblino.com/

Open source, browser-based, javascript, webgl, and can even control an Arduino version of your robot (via node); I'm not sure why it's such an underdog, but I suspect it's partially the name.

Note - a working github path is:

https://github.com/omni360/assemblino.js

Not sure why the links are broken on the page - maybe this project is "dead" - but the source lives on I guess...

Good. Dynamic datasets for next gen AI. I think simulation is going to be essential for the AI of the future: physical 3D simulation, game simulation (like AlphaGo, which run MCMC) and ultimately, complex scene with multiple human and non-human agents (self driving cars, robots, chat bots). The era of ImageNet and static data is over.

"This policy itself is still a multilayer perceptron, which has no internal state, so we believe that in some cases the agent uses its arms to store information." - That's a pretty surprising result to me!

The videos seem to have strong control flavour. What's the hook for AI researchers?

These environments are often used as a testbed for reinforcement learning, e.g. https://arxiv.org/abs/1502.05477

"However, a policy that succeeds in simulation often doesn't work when deployed on a real robot. Nevertheless, often the overall gist of what the policy does in simulation remains valid in the real world. In this paper we investigate such settings, where the sequence of states traversed in simulation remains reasonable for the real world, even if the details of the controls are not, as could be the case when the key differences lie in detailed friction, contact, mass and geometry properties"

from Transfer from Simulation to Real World through Learning Deep Inverse Dynamics Model

https://arxiv.org/abs/1610.03518

by Christiano, Shah, Mordatch, Schneider, Blackwell, Tobin, Abbeel, & Zaremba

Gait control is widely recognized as on-topic in AI robotics circles.

There are sessions on this topic at the main conferences, for example.

Well, yes, but getting robots to move without falling over is not really a core AI topic; it's an application area where some AI stuff happens to be useful.

Coming from a site with a name like "Open AI" I expected some software to help promote research on workhorse AI topics such as search, planning, constraint programming, knowledge representation and automated reasoning.

> getting robots to move without falling over is not really a core AI topic

Animals solve the problem wonderfully every day, using (at least partly) their brains. That's a working definition of an AI problem.

Robotics is not a core problem for AI. It's not even a core problem for Comp Sci. There's a whole bunch of kinematics and dynamics at play which are completely besides the point for most AI people. The only thing (some) AI folks care about is the control problems that arise in Robotics but the development of suitable methods to solve these comprises, I would argue, a rather small and not representative part of the field.

> Robotics is not a core problem for AI.

I respectfully, but firmly, disagree.

Robotics has always been at the center of the AI vision, for academics and in culture.

There's a movie called 'AI'. It's not about automatic labelling of YouTube videos or 2-sat.

Planning and reasoning more your kind of AI? Seminal AI tech STRIPS and A* were invented to drive Shakey the robot. Shakey also had very early computer vision.

Hardly anyone would claim that self-driving cars - robots with people inside - are not AI, and a pretty compelling bit of AI right now.

> I respectfully, but firmly, disagree.

Please pay attention: Robotics is just one application area for AI. Core problems from Robotics may influence the field of AI but they don't define it. Gait control or vision are or what have you are only interesting for AI people insofar as they can be used to study automated reasoning. The core problems of AI are the same as those of Computer Science: search, sort and computability.

'Please pay attention'

Hilarious. Please continue to school me on the nature and basis of artificial intelligence.

My bad. Please forgive me Mr Robot Researcher. Tell me again how all of AI exists to feed your robotic fever dreams.

>Coming from a site with a name like "Open AI" I expected some software to help promote research

Really? I'd expect some stuff that helps me integrate AI into whatever it is I'm building.

They are a privately funded research organization. It doesn't make sense to expect anything. We can just watch with interest.

This is cool. I've been working on a quadrupedal robot. I wonder if some of these tools would help me visualize and develop gaits for the robot? It looks like they would.

You could also try out pybullet, the official Python bindings for Bullet Physics which uses the same underpinnings as Roboschool: pip install pybullet and you are up and running, quickstart guide is at http://pybullet.org. It has a Minitaur quadruped robot as one of the examples.

Now, that is a minimalist website if ever I've seen one! Thank you for the recommendation.

Huge news, waiting for open source robot simulator for long. It's so wonderful. However, the shadow and appearance are a little pale compared to MuJoCo.

Theres always gazebo and a few others.

Has anyone done a comparison of this, gazebo, v-rep and mujoco?

This is really cool, been looking for something like this

This looks AMAZING. 3D games will eventually use deep learning controllers on fully simulated physical forms, instead of animations. The future is going to be amazing.

Reply on news.ycombinator.com