Training Dyna-2 at million-hour scale, repeatably — DYNA

How we rebuilt every step of our robotics data stack for million-hour scale training.

[ Research ] Training Dyna-2 at million-hour scale, repeatably Category: Research Author: Dyna Robotics Date: August 2026 Read: 22 min Robot or vendorcollectionLanding bucketobject storageTraining-ready dataMCAP episodesTraining manifestcolumnar + mmapGPU clustertraininguploadsingestioncurationloading Figure 1: Flowchart of data lifecycle From collection to training on the GPU cluster, robot data moves through four stages: collection into a landing bucket, ingestion into training-ready episodes, curation into the dataset for a given experiment, and loading into training batches.…

saved by

• Khushi Wadhwa
• Miguel Villa Floran

related reading

• Does Scaling Web-Video Pre-training Help Real Robots Do Real Work? | Rhoda AIrhoda.ai
• How to Rack 30 Petabytes of Storage | blogsi.inc
• Causal Video Models Are Data-Efficient Robot Policy Learners | Rhoda AIrhoda.ai
• Multi-Datacenter Training: OpenAI's Ambitious Plan To Beat Google's Infrastructuresemianalysis.com
• Keep the Tokens Flowing: Lessons from 16 Open-Source RL Librarieshuggingface.co
• Training great LLMs entirely from ground up in the wilderness as a startup - Yi Tayyitay.net
• How to Rack 30 Petabytes of Storage | blogsi.inc
• Nikolaus West (@NikolausWest) on Xx.com
• Network and Storage Benchmarks for LLM Training on the Cloud | Henry Zhumaknee.github.io
• how we accidentally solved robotics by watching 1 million hours of YouTube – atharva's blogksagar.bearblog.dev
• From bare metal to a 70B model: infrastructure set-up and scripts - Imbueimbue.com
• IsoCompute Playbook: Optimally Scaling Sampling Compute for RL Training of LLMscompute-optimal-rl-llm-scaling.github.io