We also have some great 3D interactive examples of our environments and rollouts on our website! We've released our simulation assets, and I'm excited to see where real2sim2real can take us 🤠 (6/N)

An amazing capability this enables is policies can learn to operate in new environments with nothing but a video of your new environment! We call this Scanned Deployed Finetuning. (5/N)

But wait, wheres the self improvement?

We realize that even 10 demos becomes a burden when you scale up to 100, 500, 1000... environments. After an initial training phase of the generalist policy, we leverage it to provide its own demos to bootstrap RL and master new environments! (4/N)

We then trained policies on these crowdsourced environments with RL bootstrapped from ~10 demos.After collecting rollouts from state-based policies, we distill them into a generalist visuo-motor policy that zero-shot/few-shot transfers to the real world! (3/N)

We crowdsourced data across 3 different continents from family & friends - anyone can do it with their phone! We use 3D reconstruction methods like Gaussian splats to make diverse, visually & geometrically realistic sim environments for training policies (2/N)

Real-world data collection is expensive and slow. How can we alleviate the effort of data collection, while allowing policies to continuously improve!?

Checkout our latest work - Robot Learning with Super Linear Scaling :)
casher-robot-learning.github.io/CASHER/

(1/N)