OpenRoboto

OpenRoboto is an open robotics-intelligence project whose Arena is a decentralized data engine for collecting and processing data for customers and robot models. The project describes its approach as “robot intelligence built in the open.”^[1]​

Overview

OpenRoboto’s public miner, protocol, and validator are provided through its Bittensor subnet, identified as netuid 80. Participants fine-tune the π0.5 base model, submit model weights on-chain, and compete on a leaderboard.^[2]​

The subnet’s software includes an evaluation harness that scores models in parallel and reproducibly across six LIBERO suites in MuJoCo.^[2] Its protocol repository contains shared contracts covering seed derivation, model fingerprinting, status vocabulary, and application programming interface (API) schemas. These are used by the backend, miner command-line interface, and evaluation worker.^[2]​

Software Tools

OpenRoboto’s command-line interface (CLI), published on the Python Package Index as OpenRoboto, provides commands for initialization, upload, submission, and status. It targets subnet netuid 80.^[2] The organization also lists a repository named jev-robot-control.^[2]​

Subnet

OpenRoboto operates as subnet 80 on the Bittensor network and focuses on evaluating and improving open-source vision-language-action (VLA) models for robotics. Participants, known as miners, fine-tune supported base models and publish complete model checkpoints to Hugging Face for evaluation. Models are assessed through simulation benchmarks and, in separate tracks, physical-robot tests, with results used to determine rankings and allocate network emissions.

The subnet uses public submission records, randomized evaluation seeds derived from Bittensor block data and the drand randomness beacon, and model verification procedures to support reproducibility and reduce manipulation. Submissions require a season-specific evaluation fee, which may be burned or transferred depending on the competition. Validators use the resulting rankings to set on-chain weights, while public APIs expose benchmark scores, leaderboards, and submission statuses. ^[3]

Key Features

• Open-Source Model Development: OpenRoboto enables participants to fine-tune vision-language-action (VLA) models and publish model weights for public access and evaluation.
• Competitive Benchmarking: Submitted models are tested in simulation against standardized robotics tasks, ranked by performance, and compared with existing models.
• On-Chain Incentives: Operating as subnet 80 on Bittensor, OpenRoboto distributes network emissions based on model performance and competition rules.
• Reproducible Evaluation: Submissions use verifiable evaluation procedures and public scoring records to support independent review and comparison.
• Real-Robot Testing: The platform includes evaluation tracks using physical robots, aiming to assess whether models can transfer their simulated performance to real-world tasks.
• Community Participation: Researchers and developers can contribute models, while miners and validators participate in model evaluation and network coordination. ^{[2] [1]}