iiinova — where your digital self keeps evolving

Nicklas Hansen @ncklashansen.bsky.social · 27/09/2026

I'm in Pittsburgh for IROS 2026 and will be presenting our work "Generalizable Robotic Insertion with World Models" tomorrow morning. Stop by our poster at G11 for a chat about world models, robot learning, and beyond! ✨

040

Nicklas Hansen @ncklashansen.bsky.social · 26/06/2026

thanks so much for the kind words, amber!

130

Nicklas Hansen @ncklashansen.bsky.social · 26/06/2026

We built a super cool interactive paper (with a live demo!) so if you want to learn more definitely check that out! Everything is open-source: code, model, dataset 🌐 Interactive paper: www.nicklashansen.com/mmbench2 📄 Paper: arxiv.org/abs/2606.27326 Work with my PhD advisor Xiaolong Wang 🙌 7/n

0170

Nicklas Hansen @ncklashansen.bsky.social · 26/06/2026

💡Importantly, we find that not all data is created equal; some data collection strategies are more effective at improving the model than others! Our curiosity approach nearly matches expert policy and human data collection while being completely autonomous. 6/n

170

Nicklas Hansen @ncklashansen.bsky.social · 26/06/2026

A simple yet effective strategy for correcting the model is *targeted* data collection. During live environment interaction, we roll out candidate trajectories in the world model, score them by predicted hallucination, and execute the most hallucination-prone one. Example: an unseen maze 5/n

180

Nicklas Hansen @ncklashansen.bsky.social · 26/06/2026

We find that hallucination in world models is, first and foremost, a *data coverage* problem. Below viz shows visited state density (blue) and our developed hallucination predictor u_r (purple) side-by-side for four different tasks. Hallucinations tend to happen in less visited regions! 4/n

181

Nicklas Hansen @ncklashansen.bsky.social · 26/06/2026

We train a 350M-parameter generative world model following the Dreamer 4 recipe: a video tokenizer + symmetric decoder is trained via masked autoencoding, and a latent dynamics model is then trained on top of frozen tokenizer latents via flow matching. 3/n

170

Nicklas Hansen @ncklashansen.bsky.social · 26/06/2026

Modern generative world models render realistic, action-controllable futures, but they also frequently *hallucinate*: generating visually fluent futures while drifting away from the ground-truth dynamics. Why? To answer this question, we need to control every part of the 🌎 training pipeline. 2/n

170

Nicklas Hansen @ncklashansen.bsky.social · 26/06/2026

Super excited to share the last paper of my PhD: "Hallucination in World Models is Predictable and Preventable" ✨ We train a 350M-parameter generative world model on a large dataset spanning 210 tasks and show that we can predict *when* hallucination will happen and use that info to fix it! 🧵1/n

15514

Nicklas Hansen @ncklashansen.bsky.social · 15/01/2026

Interested in ✨world models✨? I just open-sourced an implementation of the Dreamer 4 world model. It's in PyTorch and comes with a pretrained model + a neat little web interface that lets you interact with any of 30 DMControl tasks that I trained it on! Link: github.com/nicklashanse...

14912

Nicklas Hansen @ncklashansen.bsky.social · 10/12/2025

A good benchmark measures progress on a specific problem. I don't think it's possible to capture all aspects or applications of RL in one benchmark, so we'd first need to agree on what problem we even want to solve and then build a benchmark that acts as a (hopefully good) proxy for that problem.

211

Nicklas Hansen @ncklashansen.bsky.social · 10/12/2025

Sounds like we need to prioritize building better benchmarks

240

Nicklas Hansen @ncklashansen.bsky.social · 04/12/2025

maybe afternoon? it gets pretty cold once the sun sets

110

Nicklas Hansen @ncklashansen.bsky.social · 04/12/2025

maybe embarcadero or waterfront park? would love to join

110

Nicklas Hansen @ncklashansen.bsky.social · 30/11/2025

Yes I'd love to chat! I'll DM you

010

Nicklas Hansen @ncklashansen.bsky.social · 30/11/2025

I'm in San Diego for NeurIPS! (...I live here, but I'm also attending NeurIPS) I will be speaking at the Embodied World Models workshop on Saturday 10.00-10.30am. Let me know if you are around and would like to chat, or if you just need a food recc 🫶 (pictured: la jolla)

1121

Nicklas Hansen @ncklashansen.bsky.social · 25/11/2025

🎉 New work: “Learning Massively Multitask World Models for Continuous Control” We introduce MMBench: a 200-task RL benchmark, and Newt: a language-conditioned multitask world model trained with large-scale online RL. www.nicklashansen.com/NewtWM/ Code, checkpoints, dataset etc. are open-source!

0319

Nicklas Hansen @ncklashansen.bsky.social · 20/10/2025

I am in Honolulu for ICCV 2025 and will be speaking at the Reliable and Interactable World Models workshop at 11.20am tomorrow! Let me know if you're around and want to chat about RL, world models, or anything else :-)

140

Reposted by Nicklas Hansen

Stone Tao @stonet2000.bsky.social · 22/06/2025

The sim2real demo had some mixed success, hampered primarily by the lighting conditions of the outdoors. At least it worked sometimes! Hindsight says that despite the weather, low-cost nature, only 1 hour of training, anything working is a miracle

192

Reposted by Nicklas Hansen

Stone Tao @stonet2000.bsky.social · 28/04/2025

talk by @ncklashansen.bsky.social at the world models workshop starting now!

041

Nicklas Hansen @ncklashansen.bsky.social · 23/04/2025

That's a very valid question but unfortunately not something I'm in a position to answer as a mere PhD student :-)

010

Nicklas Hansen @ncklashansen.bsky.social · 21/04/2025

I will be in Singapore 🇸🇬 this week for ICLR 2025! Let me know if you're around and want to chat about RL, world models, LLMs, or anything else :-) I am presenting two papers on model-based RL and also speaking at the World Models @ ICLR workshop on Monday. (Photo is from my last trip to Singapore)

3121

Nicklas Hansen @ncklashansen.bsky.social · 18/04/2025

Congrats Nathan! Looking forward to reading it!

010

Nicklas Hansen @ncklashansen.bsky.social · 18/03/2025

I am at GTC 2025! DM me if you're here and would like to chat about world models, RL, or anything else :-)

050

Nicklas Hansen @ncklashansen.bsky.social · 26/02/2025

Thanks so much Mei! It was great to chat

020

Nicklas Hansen @ncklashansen.bsky.social · 21/02/2025

I finally joined 🦋! Some of you may recognize me from other sites. Here's a quick intro for new connections: 👋 I work on RL, world models, and generalization in decision-making. I'm perhaps most well known for my work on "TD-MPC2: Scalable, Robust World Models for Continuous Control" www.tdmpc2.com

Visualization of domains for which TD-MPC2 has been applied, including locomotion, manipulation, dexterous hands, humanoids, autonomous racing.

4414