# Documentation > Public page text · DVIDIA Canonical page: [Documentation](https://dvidia.org/docs) DVIDIA / Documentation # How it fits together. Open physical skills. Open-weight inference. One house, two connected paths. [The vision behind it](https://dvidia.org/manifesto) ## dvidia.org The house Human know-how, shared so more people can build with it. The skills side ### Open physical skills People demonstrate. Datasets preserve the evidence. Robots learn through training and testing. - **Demonstrations** Hands, tools, a complete task - **Reviewed datasets** Provenance, permissions, usable data - **Evaluated robot skills** Training, compatibility, real-world checks [Explore the evidence](https://dvidia.org/datasets) The open-weights desk ### OpenWeights Terminal A DVIDIA product. Research open weights, call a model, or offer a machine already running one. #### Boards Model, runtime and hardware measurements. Measurements, not a shop #### Pool Job queue, live offers and the usage ledger. 80% host · 20% desk [Inside OW Terminal](https://owterminal.com/docs) **The connection is useful intelligence.** Inference could help describe demonstrations and assist review. Training and validating a robot skill remain separate steps; this is not an automatic video-to-robot pipeline. ### Follow an inference request - **Caller gate** Virtual keys, budgets, rate limits - **OW Pool** Select a live offer; queue and bill the job - **Host** A browser tab or pool-worker claims the job - **Local engine → GPU** The host’s machine runs the model The caller gate is the separate [dvidia-inference/litellm](https://github.com/dvidia-inference/litellm) service. It forwards to the pool; it does not select the GPU. Its documentation lists no public hostname yet, so the documented direct route uses an OW pool key. A host’s local LiteLLM or engine is a different process from the caller gate. Pointing a worker back at the caller gate would loop the job. [Call or host on OW Terminal](https://owterminal.com/docs) ### What belongs in each repository? #### This DVIDIA website Capture, skillspaces, dataset discovery and the documentation you are reading. #### [Tarzelf/ow-terminal](https://github.com/Tarzelf/ow-terminal) The desk: site, boards, pool ledger, job queue and pool-worker. #### [dvidia-inference/litellm](https://github.com/dvidia-inference/litellm) The caller gateway: virtual keys, budgets and rate limits. A separate repository and process. #### Hosts’ engines Ollama, llama.cpp, LM Studio or a local proxy and engine. The model runs on the host’s GPU. This is a map of responsibilities, not one nested codebase. Repository access may require permission. ### How do open skills and paid inference coexist? Open releases keep their licenses. An inference bill pays for running a model; planned premium skillspaces pay for hosting services. Neither makes a physical skill exclusive. OW’s 80/20 split applies to its completed-job ledger: 80% owed to the host, 20% recorded for the desk. That split uses the host’s pool offer; a caller gateway can bill a different price. It is not a robotics contributor royalty or an automatic transfer into a skills fund. Funding and payouts require desk settlement; a queued payout is not a sent payment. DVIDIA’s current OW integration is a read-only research catalog. Routing video analysis through the pool is a future integration, not a connection this diagram activates. Start with the skills side ## What would you like to do? One useful demonstration is a place to start. From evidence to ability ## A video is the beginning. - Demonstration - Reviewed dataset - Trained & tested skill Each step needs its own validation. A recording alone is not a working robot policy. ### Where do my recordings go? Browser camera recordings and saved drafts stay in this browser. Saving a draft does not upload or publish it. Download original videos you want to keep. Uploads and contributions use separate, explicit consent. Camera access needs permission and a secure connection; device support varies. Illustrative footage and clip counts do not establish robot success. ### Can I export a robot training dataset? In development Native datasets target LeRobot v3. Current recording downloads are original video files, not validated training packs. A valid pack needs frame data in Parquet, playable camera media, complete metadata and consistent episode offsets. The release check is loading and replaying the real frames with the upstream LeRobot reader. [LeRobot v3 documentation](https://huggingface.co/docs/lerobot/lerobot-dataset-v3) ### Is there a robot installer? In development The separate CLI repository and this app’s in-tree CLI differ. Both are private previews, not verified general-purpose robot installers. Consolidation, complete artifact transfer and upstream validation must pass before an installation command is published. [CLI repository](https://github.com/Dvidia-Inference/dvidia-cli-archived) ### What stays open? Underlying releases stay accessible under their licenses. Paid workspace services and contributor revenue sharing are planned; rates and terms must be published before charging. Check third-party licenses at their original sources. [OW_ **Looking for open-weight inference?** OW Terminal is DVIDIA’s open-weights desk.](https://owterminal.com) Inference endpoints and pool accounting belong to OW Terminal. This DVIDIA website does not run a second inference pool. --- [Agent index](https://dvidia.org/llms.txt) · [Markdown mirror](https://dvidia.org/docs/mirror.md)