- Hard Startups
- Robotic Legos, Autonomous Jets, and Space Delivery
Robotic Legos, Autonomous Jets, and Space Delivery
Hey y’all in this issue we have:
- Team at Standford teach a robot to build legos
- Anduril unveils autonomous jet capable of Mach 0.95 speeds
- Space based package delivery company tests recovery system
AI is coming for your hobbies. Some folks at Stanford taught a robot how to build lego structures in simulation and transferred those skills to a real robot with no additional training
They released their paper along with some incredible videos on the project page this week (definitely watch the videos to get the full effect. pretty wild to see a robot hand do a complicated task like this automatically).
Teaching robots to be good at multi-step tasks
We’re already pretty good at teaching robots to do specific single tasks, but tackling problems that require multiple actions over time (like locating, grabbing, and placing a lego block in the correct target position) is still hard.
Part of the reason it’s hard is that a subsequent action’s success rate could depend on how an earlier action was completed. For example, if a robot grips the lego in an orientation that’s not optimal for inserting the lego, then the insertion task is more likely to fail.
The paper explores a new way to train a robot across a whole sequence of actions. They start with a sequence of distinct actions (Search, Orient, Grasp, Insert) and introduce a novel way to learn a policy for each action that maximizes the chance of completing the bigger picture task.
With this type of training, it’s more likely that each action is performed in a way that makes the subsequent actions more likely to complete.
The team trained the policies on 1024 simulated robot agents using NVIDIA’s Isaac Gym. After training was complete they were able to take the skills and transfer them to a real robot with no additional training.
Sim-to-real training seems to be getting more popular as simulation environments like Isaac Gym are starting to mature and the tools are becoming easier to use for researchers.
Anduril buys an autonomous jet company
Anduril announced their Fury jet platform this week. They acquired the jet from a North Carolina based company Blue Force Technologies, which had been working on the Fury vehicle since 2019.
The jet is capable of Mach 0.95 speeds and has no onboard human pilot. The lack of human pilot makes the production and operation of this type of aircraft significantly cheaper.
The announcement even generated some interesting back and forth from Elon Musk and Palmer Luckey (founder of Anduril).
Cargo Delivery from space
Inversion Space is building the fastest way to send cargo across the earth, through space!
The YC-backed (S21) company posted a video on LinkedIn this week of a successful test of their recovery system on their test vehicle Ray.
Their system could one day deliver cargo anywhere on Earth in under an hour, and provide back and forth resupply for the growing space industry in LEO.
That’s all for today!
Let me know what you thought, just reply back to this email with any feedback