Benedikt Stroebl
@benediktstroebl
Followers
798
Following
3K
Media
31
Statuses
350
co-founder and cto @ludus_labs (YC S25), ex-Princeton, Oxford, Google, TUM
sf
Joined November 2020
I am leaving Princeton to start @ludus_labs with Venia and Gianluca. We’re generalizing AI agent evaluations into the physical—making the frontier of intelligence something everyone can watch. The first research-first entertainment company. Stay tuned.
11
9
92
congrats! great to see this.
Today, I’m launching a deeply personal project. I’m betting $100M that we can help computer scientists create more upside impact for humanity. Built for and by researchers, including @JeffDean & @jpineau1 on the board, @LaudeInstitute catalyzes research with real-world impact.
0
0
2
This Saturday: come and hack autonomous fighting roombas with us in sf! Yes, vacuums. We provide roombas and everything else to train your first fighter. Getting started is simple. RSVP below.
I’m excited to release our repository for simulating, training, and deploying (real) jailbroken Roombas that can fight. We replicated our living room octagon in simulation and can train autonomous fighting Roombas under different sensor setups using RL (see video). (1/4)
2
1
14
At Ludus, we are convinced that sports offer a great testbed for standardized and dynamic robotics evals. In that sense we are building on a rich history from RoboCup etc.
The evaluation problem in robotics differs from other ML fields because of (1) the online nature of real-world evals, and (2) a dynamic real world that makes standardizing evaluation tasks across labs and across time difficult.
0
0
7
Everyone training control policies knows how brittle it can be starting from scratch; especially with humanoids. Improving the adaptability of pretrained BFMs makes this much easier!.
Behavioral Foundation Models (BFMs) trained with RL are secretly more powerful than we think. BFM’s directly output a policy believed to be near-optimal given any reward function. Our new work shows that they can actually do much better:
0
0
6
The roomba getting some last tips before its big fight next Saturday!.
0
0
4
Looking for a free robot?. Come watch or compete in the first ever AI roomba fight next Saturday (28th). Programmer to champion. All you need is RL. Champ walks away with a roomba. Chaos guaranteed.
1
0
7
RT @ludus_labs: You are cordially invited to the first ever ai roomba fight. Watch gladiator roombas or join the hackathon and train your….
0
5
0
RT @benediktstroebl: How to build an agent that works? A lot can go wrong. This write-up echoes many experiences from working on agents for….
0
1
0
How to build an agent that works? A lot can go wrong. This write-up echoes many experiences from working on agents for a while now. - agent frameworks are surprisingly brittle .- don’t succumb to building (overly complex) multi-agent scaffolds.- share more context. The post.
I see a lot of people make the same mistakes building agents. So we shared a few of the principles we use.
0
1
3
RT @gianlucabencomo: Lots of people are working on robotic hands, but almost no one is working on robotic feet — even though humanoid locom….
0
1
0
1
0
0
RT @ryanmart3n: Announcing OpenThinker3-7B, the new SOTA open-data 7B reasoning model: improving over DeepSeek-R1-Distill-Qwen-7B by 33% on….
0
190
0