Does Scaling Web-Video Pre-training Help Real Robots Do Real Work? | Rhoda AI

We investigate scaling model size and pre-training compute for video-based robot policies, and benchmark on a real industrial manipulation task.

Video pre-training has recently become a cornerstone of robot foundation models, and with it, a widely held belief: scaling web-video pre-training leads to better downstream robot task performance. This belief has not been rigorously tested. Few studies examine the effect of scaling model size and compute, instead often substituting a binary comparison against a model trained from scratch. Fewer still have stepped beyond proxy metrics and simple lab tasks to investigate scaling on complex, long-horizon tasks representative of real-world use cases. We put this belief to the test with Direct…

saved by

• Ishaan Panigrahi
• Alex K
• Miguel Villa Floran

related reading

• Causal Video Models Are Data-Efficient Robot Policy Learners | Rhoda AIrhoda.ai
• RL Scaling Laws for LLMs - by Cameron R. Wolfe, Ph.D.cameronrwolfe.substack.com
• Training Dyna-2 at million-hour scale, repeatably — DYNAdyna.co
• Pretraining progress is mostly coming from datadwarkesh.com
• Frontier of General-Purpose Robotics (2025)adampatni.com
• Emergence of Human to Robot Transfer in Vision-Language-Action Modelspi.website
• Generalist - GEN-0 / Embodied Foundation Models That Scale with Physical Interactiongeneralistai.com
• How Claude Performs on Robotics Tasks \ Anthropicanthropic.com
• [2510.13786] The Art of Scaling Reinforcement Learning Compute for LLMsarxiv.org
• Real-Time Action Chunking with Large Modelspi.website
• IsoCompute Playbook: Optimally Scaling Sampling Compute for RL Training of LLMscompute-optimal-rl-llm-scaling.github.io
• How Can We Make Robotics More like Generative Modeling? | Eric Jangevjang.com