3D representations are more critical than what we thought for manipulation. Our work 3D Diffuser Actor marries those with policy diffusion and achieves a new SOTA on both RLBench and CALVIN! 3d-diffuser-actor.github.io 🦾 with Tsung-Wei Tsung-Wei Ke and Katerina Katerina Fragkiadaki

Fantastic gang from TU Darmstadt Intelligent Autonomous Systems Group Georgia Chalvatzaki and offspring at Shmoopy …

Tired of labeling your robot data?🤖 Excited to present NILS, a novel approach that leverages foundation models 🧠 to segments videos into tasks and generates semantically meaningful natural language annotations 💬 of varying levels of granularity! Web: robottasklabeling.github.io

Turn a single pre-trained model’s layers into MoE “experts” and reuse them? Finetuning a “router” slightly cuts loss—cool proof of concept. Can we combine dynamic compute paths/reuse + coconut-like latent reasoning? jyopari.github.io/posts/reuse

⚡️ Do you think training robot locomotion needs large scale simulation? Think again! Our new paper shows how to train an omnidirectional locomotion policy directly on a real quadruped robot in just a few minutes 🚀 Top speeds of 0.85 m/s, two different control approaches, indoor

Happy to share to be a recipient of the 2025 Apple Scholars in AI/ML PhD Fellowship Programm! Grateful to @apple for this recognition and the support. Thank you to all my collaborators, colleagues and to supervisor who made this possible! machinelearning.apple.com/updates/apple-…

Llama 4 (Meta) shows too much SFT limits RL exploration — something we also found in our recent work! A new and superior pretraining paradigm is around the corner to unleash a new era of reasoning. Check out our paper: arxiv.org/abs/2502.19402 Thread: x.com/pulkitology/st…

⚡️ Can one policy control 1000 different robots? 🤖 We explore Embodiment Scaling Laws: Training on more diverse robot embodiments boosts generalization 📈 Our generalist policy, trained on 1000 generated robots, zero-shot transfers to the real Go2 quadruped and H1 humanoid 🚀

Imitation learning has a data scarcity problem. Introducing EgoDex from Apple, the largest and most diverse dataset of dexterous human manipulation to date — 829 hours of egocentric video + paired 3D hand poses across 194 tasks. Now on arxiv: arxiv.org/abs/2505.11709 (1/4)

📢📢📢 Neural Inverse Rendering from Propagating Light 💡 Our CVPR Oral introduces the first method for multiview neural inverse rendering from videos of propagating light, unlocking applications such as relighting light propagation videos, geometry estimation, or light

What if an LLM could update its own weights? Meet SEAL🦭: a framework where LLMs generate their own training data (self-edits) to update their weights in response to new inputs. Self-editing is learned via RL, using the updated model’s downstream performance as reward.