World models for policy evaluation; the fact that world model performance is highly correlated with real-world performance is incredibly valuable on its own.

Interesting conversation - I am particularly curious about how WMs/video models stand to benefit from object centric interaction data with just enough causal and semantic descriptors

Love seeing egocentric data treated as more than "just video". Explicit state transitions (hands, objects, actions) is the missing piece between raw videos and world models. Curious to see how these scale across tasks and environments!

Sanskar Pandey already published a 1h PoC enrichment of Egocentric-100K dataset sanskxr02/fidelity on HF x.com/sanskxr02/stat…

Today we’re open-sourcing a small first subset of state–action–state′ trajectories from egocentric factory video. Built on Build AI data and enriched by us, this is early infrastructure for learning-ready video in world models and offline RL. Still very much iterating on the

We believe the next breakthrough in long-horizon agents is training models to manage their own context. Introducing our new research direction on Recursive Language Models. We are sharing our initial experiments showing the promise of RLMs. primeintellect.ai/blog/rlm

Teleoperation data does not seem to scale well on its own -- we need lots of video data of humans and robots interacting with their environment to learn physics and environment interaction. This will unlock general robot intelligence.

This is an interesting take for sure,the biggest issue I’ve had with any “in-the-wild” data collection technique is simply that the data collected is constrained to simple grasping or quasi-static pick-and-place actions

To drive the point home - regardless of how detailed your IMU collection stack is, this point still holds. Trying to recover n-DoF actions does not solve the embodiment gap. It really just seems like a painful way of solving for an under-specified inverse problem