Trevor McInroe

Research Scientist @SkildAI.

PhD, The University of Edinburgh.

Interested in intelligence through trial-and-error in the Era of Experience.

[Trevor McInroe @trevormcinroe]

Introducing Terra Nova, a new comprehensive challenge environment for RL research inspired by Civilization V.

[Trevor McInroe @trevormcinroe]

Feels like a good time to announce that I've joined Skild as a Research Scientist.

Looking forward to continuing the push for RL in the real world.

[Skild AI @SkildAI]

We have acquired Zebra Technologies’ robotics arm (formerly Fetch Robotics).

This is what happens when orchestration meets intelligence -- a major step toward fully autonomous warehouses.

More robots. More environments. One unified brain.

[Trevor McInroe @trevormcinroe]

@dirtman Are you sure quantity is the best metric for comparison?

[Trevor McInroe @trevormcinroe]

@Sentdex Try MBPO+SAC. Then you can explore within the imagined trajectories of the model with a variety of strategies.

Also can cut SAC's entropy target in half, e.g., -dim(A)/2

[Trevor McInroe @trevormcinroe]

@arnie_hacker This presupposes a convergence on the "correct" morphology.

[Trevor McInroe @trevormcinroe]

@alpercanbe re (2) This is harder to nail down and likely depends on audience. You're right that we usually don't write out xent loss, but most RL papers still have the obligatory paragraph defining the MDP and learning objective "\pi^{\ast} = \argmax_{\pi} \mathbb{E}_{\pi} \sum_t ..." ;)

[Trevor McInroe @trevormcinroe]

@alpercanbe It's a mixture of (1) relevancy and (2) vibes.

re (1) If I am writing a paper on optimizers that improve upon Adam, I'll need to explicitly write out the maths of Adam. If the paper is on something else and I'm just using Adam, I can just cite (Kingma & Ba, 2015).

[Trevor McInroe @trevormcinroe]

@ID_AA_Carmack There's been some work on the disconnect between RL and other areas of ML w.r.t. NN size. Check out https://t.co/g0NDdL8SFA and https://t.co/EI5LXUH3oO

[Trevor McInroe @trevormcinroe]

@kevin_zakka Keep up the great work.

[Trevor McInroe @trevormcinroe]

@willccbb @TheZachMueller Return to monke

[Trevor McInroe @trevormcinroe]

@yoavgo An (usually learned) approximation of the MDP's transition function and reward function.

In the POMDP case, we're likely modeling the observation function instead of the state transition-function directly.

[Trevor McInroe @trevormcinroe]

@thegautamkamath @TmlrOrg Living in the UK these past four years has made me miss Mexican food to an extreme degree...

[Trevor McInroe @trevormcinroe]

@Stone_Tao Thanks!

I've been using ManiSkill for a world-modeling project lately. Now that's cool stuff

[Trevor McInroe @trevormcinroe]

Introducing Terra Nova, a new comprehensive challenge environment for RL research inspired by Civilization V.

[Trevor McInroe @trevormcinroe]

Music credit:
"A Brighter Future" by All Good Folks: https://t.co/QipaQQ7GZY
"Go Ahead" by Gerald Olivieri: https://t.co/T42eNalRWg

[Trevor McInroe @trevormcinroe]

On a completely unrelated note, if anyone wants to play a game of Civ, feel free to DM me!