Trevor McInroe (@trevormcinroe) header
Trevor McInroe (@trevormcinroe) profile picture

Trevor McInroe

@trevormcinroe

Research Scientist @SkildAI. PhD in RL, The University of Edinburgh. Interested in intelligence through trial-and-error in the Era of Experience.

 San Francisco, CA
trevormcinroe.github.io
Joined September 2021
  • Tweets101
  • Following439
  • Followers156
7Media
Photo by Trevor McInroe (@trevormcinroe)Photo by Trevor McInroe (@trevormcinroe)Photo by Trevor McInroe (@trevormcinroe)Photo by Trevor McInroe (@trevormcinroe)Photo by Trevor McInroe (@trevormcinroe)Photo by Trevor McInroe (@trevormcinroe)Photo by Trevor McInroe (@trevormcinroe)
Introducing Terra Nova, a new comprehensive challenge environment for RL research inspired by Civilization V.
4
6
6
49
5,382
The humanoids just want to (self) play.
Skild AI (@SkildAI) profile pictureSkild AI
@SkildAISep 23
We trained a robot to play football. How? Self-play for 140 years in a virtual World Cup. Meet the #Messinator:
5
217
Trevor McInroe retweeted
Skild AI (@SkildAI) profile picture
Skild AI
@SkildAISep 10
We just hit 100M ARR within 10 months of starting deployments. We are in factory lines. On construction sites. In kitchens. In data centers. Cleaning. Welding. Building. Cooking. Deploying.
120
291
163
2,148
1,687,974
Some exciting new results with in-context learning, allowing for very long-horizon behaviors. The team hit it out of the park with this one!
Skild AI (@SkildAI) profile pictureSkild AI
@SkildAIAug 25
Introducing S1, our new foundation model that learns from one example. It can be taught 10-minute long tasks that it has never seen before, from one video prompt without any fine-tuning. Watch S1 operate in real-time via in-context learning:
1
18
364
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) profile pictureSkild AI
@SkildAIApr 15
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.
10
301
On a completely unrelated note, if anyone wants to play a game of Civ, feel free to DM me!
1
4
360
Music credit: "A Brighter Future" by All Good Folks: uppbeat.io/t/all-good-folks/… "Go Ahead" by Gerald Olivieri: uppbeat.io/t/gerald-olivieri…
3
303
2
5
569
The ability to feel will be crucial for real-world robots. Our new NeurIPS paper studies how tactile sensors affect downstream RL control. With the right representations, we can do dexterous tasks like Baoding ball rotation without seeing anything at all (no pixels, just touch).
Elle Miller (@ElleAMiller) profile pictureElle Miller@ElleAMiller17 Nov 2025
Excited to share our #NeurIPS2025 paper "Enhancing Tactile-based RL for Robotic Control"! We show how incorporating self-supervision into tactile RL can lead to simulated superhuman dexterity🤏 By the way, this agent is fully blind. Find out more: elle-miller.github.io/tactil… 🧵👇
1
261
I've not done a great job promoting my work during the PhD, so here's one from last year: Given an interaction budget, what data should we collect online after pretraining an RL agent offline? We answered this question by planning to go out of distribution (PTGOOD).
1
1
430
more replies
We find that PTGOOD consistently produces the best-performing policies among a set of OtO and exploration algorithms across various continuous-control tasks and offline dataset qualities.
Image from Trevor McInroe (@trevormcinroe)
1
147
We also investigate the failure modes and complications of common exploration approaches in online RL, performance impacts of hyperparameter selection, perform comprehensive tuning of our baselines, and more. Check out the paper here: arxiv.org/abs/2310.05723
108
Load more