RoboTTT: Context Scaling for Robot Policies

Yunfan Jiang, Yevgen Chebotar, Ruijie Zheng et al.1 min readarXivJul 16, 2026

Key Takeaway for Students

Recent robot foundation models operate with single-step or short-history visuomotor context. We introduce Test-Time-Training Robot Policies (RoboTTT), a robot model and training recipe that scale visuomotor context to 8K timesteps, three orders of magnitude beyond state-of-the-art policies, without growing inference latency. At this context length, we unlock new robot capabilit

87% overstate-of-the-art policies, without growing inference latencyoutperforms the same model pretrained with 1K timesteps by 62%, sugges

Reading in Plain English mode — toggle below to switch

Recent robot foundation models operate with single-step or short-history visuomotor context. We introduce Test-Time-Training Robot Policies (RoboTTT), a robot model and training recipe that scale visuomotor context to 8K timesteps, three orders of magnitude beyond state-of-the-art policies, without growing inference latency. At this context length, we unlock new robot capabilit

This research paper advances Multimodal and production ML, directly affecting how engineers build, optimize, and ship AI systems. Teams running production LLM workloads need to understand these techniques to stay competitive in hiring and architecture decisions.

Read original on arXiv

Career Insight

medium demand

Why it matters

Skills to learn

vision encoder integrationcross-attention in VLMsmultimodal data pipelines

Interview questions

How would you design a system that applies the core technique from "RoboTTT" at 10k requests/sec?
What are the production trade-offs of adopting Multimodal in a latency-sensitive serving stack?
Walk me through the key architectural decision in "RoboTTT" and when you would NOT use this approach.

Relevant for ~12 months · generated by the enrichment pipeline