With the amount of new subnets being added it can be hard to get up to date information across all subnets, so data may be slightly out of date from time to time
With the amount of new subnets being added it can be hard to get up to date information across all subnets, so data may be slightly out of date from time to time
Kinitro is Bittensor’s Subnet 26, a project focused on advancing robotic intelligence through incentivized competitions. (This subnet was previously known as Storb, which centered on decentralized storage, before being sold and rebranded as Kinitro in mid-2025.) In its new form, Kinitro creates a decentralized competition platform where AI researchers and developers train embodied AI agents (robots or simulation agents) to perform tasks, and are rewarded based on performance. The aim is to accelerate the development of embodied intelligence – AI that can understand and act in the physical world – by tapping into Bittensor’s decentralized network of miners and validators.
In simpler terms, Kinitro turns AI training into a competition. Network participants called validators propose specific tasks or challenges (for example, a robotic navigation or planning problem) along with how success is measured and what rewards are offered. Miners (developers) then build and train AI agents to tackle these tasks and submit their trained models as entries. The validators evaluate each submitted agent and automatically dispense rewards to the best-performing ones. This process incentivizes continuous improvement: better-performing agents earn more, encouraging miners to keep enhancing their models, while validators earn rewards for providing accurate and timely evaluations.
Kinitro is Bittensor’s Subnet 26, a project focused on advancing robotic intelligence through incentivized competitions. (This subnet was previously known as Storb, which centered on decentralized storage, before being sold and rebranded as Kinitro in mid-2025.) In its new form, Kinitro creates a decentralized competition platform where AI researchers and developers train embodied AI agents (robots or simulation agents) to perform tasks, and are rewarded based on performance. The aim is to accelerate the development of embodied intelligence – AI that can understand and act in the physical world – by tapping into Bittensor’s decentralized network of miners and validators.
In simpler terms, Kinitro turns AI training into a competition. Network participants called validators propose specific tasks or challenges (for example, a robotic navigation or planning problem) along with how success is measured and what rewards are offered. Miners (developers) then build and train AI agents to tackle these tasks and submit their trained models as entries. The validators evaluate each submitted agent and automatically dispense rewards to the best-performing ones. This process incentivizes continuous improvement: better-performing agents earn more, encouraging miners to keep enhancing their models, while validators earn rewards for providing accurate and timely evaluations.
To outline the process more concretely, Kinitro’s competition workflow involves three key steps:
By structuring AI development as above, Kinitro “drives the future of robotic policy and planning models with incentivized competitions”. This directed incentive model (the project’s motto is cerebrum machinae, meaning “mind of the machine”) is intended to push the frontier of robotics AI. Over time, tasks can become more complex or change (“benchmarks evolve”), and miners must adapt their agents, ensuring a continuous improvement cycle in the collective embodied AI capability.
Kinitro’s product is essentially a decentralized platform for AI competitions built on the Bittensor blockchain. Technically, it consists of a few core components working together: a backend platform, a network of validator nodes, and participating miner nodes (which run the AI agents). The backend and blockchain logic orchestrate the posting of tasks, submission of models, and distribution of rewards. Validators are responsible for hosting and evaluating tasks (for example, running a robot simulation or test environment to judge an agent’s performance), while miners focus on developing the agents. All of this is underpinned by Bittensor’s substrate-based network, meaning the coordination and rewards are handled in a decentralized manner (with Kinitro’s own α (alpha) subnet token as the unit of reward).
The technical architecture of Kinitro is actively evolving. Initially, the developers experimented with a “parent-child validator” design – possibly where a primary validator could spawn subtasks – but they decided to refactor this into a clearer platform + validator architecture. In the refactored architecture, a dedicated competition backend (platform) coordinates multiple validators more directly, improving scalability and clarity. Each validator can independently evaluate agents and report results back to the platform, rather than one validator hierarchically handling all subtasks. This modular approach makes it easier to add or remove validators and ensures the system can handle many simultaneous competitions.
Under the hood, Kinitro’s codebase is written mostly in Python. It likely leverages existing Bittensor frameworks for networking and consensus, while adding new logic for managing competitions. For example, smart contracts or on-chain logic may be used to record when a task is posted, which miners have submitted models, and how rewards are allocated, ensuring transparency and trustlessness. Off-chain, there may be integration with AI frameworks or simulators to actually run the robots’ evaluations (since training and testing AI agents involves significant compute). Kinitro’s repository provides a “Kinitro agent template” for miners to start building their agents, indicating that there are defined APIs or guidelines for how an agent should be packaged and submitted to the network.
The incentive structure of the build is an important aspect: Kinitro’s blockchain logic automatically mints and distributes α rewards based on performance. Validators earn emission rewards for doing the work of evaluation (and doing it accurately and promptly), while miners earn the posted task bounties in proportion to how well their agent did. This creates a balanced economy where both creating good AI models and honestly assessing them are rewarded. All transactions and rewards occur on-chain in a decentralized fashion.
From a security and infrastructure standpoint, the team is working on hardening the system. For instance, secure communication between the backend platform and validators is being improved to prevent tampering or cheating in the evaluation process. They are also integrating monitoring and telemetry tools (like PostHog) to track the performance of the network and detect any issues in real-time. Logging has been identified as an area to improve as well, so that the behavior of agents and validators can be audited if needed. All these technical enhancements are part of ensuring that Kinitro’s platform is robust and enterprise-grade as it moves from prototype toward production.
In summary, the “product” of Kinitro is a full-stack decentralized AI competition system. On the front end (conceptually) it looks like a leaderboard of AI agents competing to solve robotic tasks. On the back end, it is a combination of blockchain smart contracts, networking infrastructure, and AI evaluation pipelines. This build allows anyone with the skills to become a miner (AI trainer) or a validator (competition host) by running the Kinitro code, thereby decentralizing the advancement of robotics AI beyond any single lab or company.
To outline the process more concretely, Kinitro’s competition workflow involves three key steps:
By structuring AI development as above, Kinitro “drives the future of robotic policy and planning models with incentivized competitions”. This directed incentive model (the project’s motto is cerebrum machinae, meaning “mind of the machine”) is intended to push the frontier of robotics AI. Over time, tasks can become more complex or change (“benchmarks evolve”), and miners must adapt their agents, ensuring a continuous improvement cycle in the collective embodied AI capability.
Kinitro’s product is essentially a decentralized platform for AI competitions built on the Bittensor blockchain. Technically, it consists of a few core components working together: a backend platform, a network of validator nodes, and participating miner nodes (which run the AI agents). The backend and blockchain logic orchestrate the posting of tasks, submission of models, and distribution of rewards. Validators are responsible for hosting and evaluating tasks (for example, running a robot simulation or test environment to judge an agent’s performance), while miners focus on developing the agents. All of this is underpinned by Bittensor’s substrate-based network, meaning the coordination and rewards are handled in a decentralized manner (with Kinitro’s own α (alpha) subnet token as the unit of reward).
The technical architecture of Kinitro is actively evolving. Initially, the developers experimented with a “parent-child validator” design – possibly where a primary validator could spawn subtasks – but they decided to refactor this into a clearer platform + validator architecture. In the refactored architecture, a dedicated competition backend (platform) coordinates multiple validators more directly, improving scalability and clarity. Each validator can independently evaluate agents and report results back to the platform, rather than one validator hierarchically handling all subtasks. This modular approach makes it easier to add or remove validators and ensures the system can handle many simultaneous competitions.
Under the hood, Kinitro’s codebase is written mostly in Python. It likely leverages existing Bittensor frameworks for networking and consensus, while adding new logic for managing competitions. For example, smart contracts or on-chain logic may be used to record when a task is posted, which miners have submitted models, and how rewards are allocated, ensuring transparency and trustlessness. Off-chain, there may be integration with AI frameworks or simulators to actually run the robots’ evaluations (since training and testing AI agents involves significant compute). Kinitro’s repository provides a “Kinitro agent template” for miners to start building their agents, indicating that there are defined APIs or guidelines for how an agent should be packaged and submitted to the network.
The incentive structure of the build is an important aspect: Kinitro’s blockchain logic automatically mints and distributes α rewards based on performance. Validators earn emission rewards for doing the work of evaluation (and doing it accurately and promptly), while miners earn the posted task bounties in proportion to how well their agent did. This creates a balanced economy where both creating good AI models and honestly assessing them are rewarded. All transactions and rewards occur on-chain in a decentralized fashion.
From a security and infrastructure standpoint, the team is working on hardening the system. For instance, secure communication between the backend platform and validators is being improved to prevent tampering or cheating in the evaluation process. They are also integrating monitoring and telemetry tools (like PostHog) to track the performance of the network and detect any issues in real-time. Logging has been identified as an area to improve as well, so that the behavior of agents and validators can be audited if needed. All these technical enhancements are part of ensuring that Kinitro’s platform is robust and enterprise-grade as it moves from prototype toward production.
In summary, the “product” of Kinitro is a full-stack decentralized AI competition system. On the front end (conceptually) it looks like a leaderboard of AI agents competing to solve robotic tasks. On the back end, it is a combination of blockchain smart contracts, networking infrastructure, and AI evaluation pipelines. This build allows anyone with the skills to become a miner (AI trainer) or a validator (competition host) by running the Kinitro code, thereby decentralizing the advancement of robotics AI beyond any single lab or company.
Kinitro is developed and maintained by a small, specialized team under the banner of ThreeTau. ThreeTau is the entity that took over Subnet 26 and pivoted it to the Kinitro project. The open-source repository lists three core contributors to Kinitro’s code: Ray Okamoto, Syeam Bin Abdullah, and Rishi.
Ray Okamoto – Ray appears to be one of the lead developers; he opened several of the project’s key issues and features (including documentation and security improvements) which suggests he is deeply involved in the architecture and direction of the project.
Syeam Bin Abdullah (GitHub handle “Shr1ftyy”) – Syeam is another principal engineer on Kinitro. He has contributed major code refactors (for example, overhauling the validator architecture and integrating the competition system). His involvement indicates expertise in backend/platform development for the subnet.
Rishi (GitHub handle “rishiad”) – Rishi is also listed as a contributor and presumably provides development support or research expertise to the project.
Together, this team is building Kinitro in the open via the ThreeTau GitHub organization. The ThreeTau name and the presence of a domain (threetau.net) suggest it might be a small startup or collective formed around building Bittensor subnets, with Kinitro being a flagship project. The team actively engages with the Bittensor community; for example, they communicate progress updates and calls for miners on X (Twitter), and they coordinate with the broader Bittensor developer ecosystem (e.g., via Discord, as noted in their README).
It’s worth noting that prior to ThreeTau’s involvement, Subnet 26’s previous project (Storb) had a different team focused on storage. With the transition to Kinitro, the ThreeTau team has effectively assumed ownership of this subnet. This change in leadership brought a new vision (robotics competitions) and the technical expertise needed to implement it. The current team’s mix of systems programming skills and AI knowledge is well-suited for Kinitro’s ambitious goal of merging blockchain incentives with robotics training.
Kinitro is developed and maintained by a small, specialized team under the banner of ThreeTau. ThreeTau is the entity that took over Subnet 26 and pivoted it to the Kinitro project. The open-source repository lists three core contributors to Kinitro’s code: Ray Okamoto, Syeam Bin Abdullah, and Rishi.
Ray Okamoto – Ray appears to be one of the lead developers; he opened several of the project’s key issues and features (including documentation and security improvements) which suggests he is deeply involved in the architecture and direction of the project.
Syeam Bin Abdullah (GitHub handle “Shr1ftyy”) – Syeam is another principal engineer on Kinitro. He has contributed major code refactors (for example, overhauling the validator architecture and integrating the competition system). His involvement indicates expertise in backend/platform development for the subnet.
Rishi (GitHub handle “rishiad”) – Rishi is also listed as a contributor and presumably provides development support or research expertise to the project.
Together, this team is building Kinitro in the open via the ThreeTau GitHub organization. The ThreeTau name and the presence of a domain (threetau.net) suggest it might be a small startup or collective formed around building Bittensor subnets, with Kinitro being a flagship project. The team actively engages with the Bittensor community; for example, they communicate progress updates and calls for miners on X (Twitter), and they coordinate with the broader Bittensor developer ecosystem (e.g., via Discord, as noted in their README).
It’s worth noting that prior to ThreeTau’s involvement, Subnet 26’s previous project (Storb) had a different team focused on storage. With the transition to Kinitro, the ThreeTau team has effectively assumed ownership of this subnet. This change in leadership brought a new vision (robotics competitions) and the technical expertise needed to implement it. The current team’s mix of systems programming skills and AI knowledge is well-suited for Kinitro’s ambitious goal of merging blockchain incentives with robotics training.
Kinitro is currently in an early launch phase, with the team actively working towards a full mainnet deployment of its competition system. As of late August 2025, the developers indicated they are “onboarding miners very soon” and that the validator code (for running tasks) is in the process of coming online. In fact, a project update hinted that they were targeting early September 2025 for mainnet readiness, meaning the first live competitions and reward distributions on the subnet are imminent.
Key items on Kinitro’s roadmap include:
Mainnet Competition Launch: Finalizing the implementation of the competition marketplace and evaluation system. A GitHub issue titled “Competitions system – Eval Marketplace” was opened in late August 2025, highlighting this as a feature in progress. This involves enabling validators to create tasks on-chain and a mechanism for miners to upload their models, with all participants able to view ongoing contests and their outcomes. The inclusion of a robust scoring system for evaluating agent performance is part of this launch milestone (an issue “Incorporate scoring” is being tracked as well).
Security and Reliability Enhancements: Before (and continuing after) launch, the team is hardening the platform. For example, they plan to improve logging for better debug and audit capabilities and to integrate telemetry/monitoring (using tools like PostHog) to get real-time alerts on the system’s health. Additionally, ensuring secure communications between components is on the roadmap – an issue to “harden security of backend-validator communications” is open, indicating that protecting data integrity during agent evaluation is a priority (this prevents miners from cheating or falsifying results, for instance).
Refactoring and Optimization: The team is iterating on the architecture for efficiency. The shift from a parent-child validator model to a refined platform/validator separation is one such refactor underway. These changes aim to optimize how tasks are handled and scaled – for example, to allow multiple validators to jointly handle a large competition or to make it easier to add new validators without downtime. Optimizations may also target the performance of running AI evaluations (since robotic simulations can be compute-intensive, any improvements in how the network distributes this load will be beneficial).
Documentation and Miner Tooling: Recognizing the need to attract participants, the roadmap includes creating comprehensive documentation and tutorials. There are tasks open to “add docs to [the] website” and to provide initial documentation for users. This will likely cover how new miners can create an agent (using the provided template) and join competitions, how validators can set up a node and post challenges, and general guidelines for the formats/protocols used. Improved tooling (perhaps command-line interfaces or dashboards for monitoring competitions) may also be part of the user-facing roadmap as Kinitro matures.
Future Competitions and Benchmarks: While not explicitly detailed in the issues, one can infer that once the platform is stable, a pipeline of challenge tasks (benchmarks) will be rolled out. These could range from simple tasks (like maneuvering a virtual robot through a maze) to more complex ones (multi-agent coordination, or real-world data tasks) as the community grows. The roadmap likely envisions regular competition rounds that will continuously push the state-of-the-art of the agents on the subnet. Each competition’s learnings might feed into the design of the next, creating an evolving benchmark suite (as hinted by the “continuous improvement” philosophy).
In summary, the near-term roadmap is about getting Kinitro to a fully functional state on mainnet: launching the first incentivized robotic AI challenges, ensuring the system is secure and stable, and bringing in participants. Beyond that, the project’s success will be measured by its ability to attract AI talent to continuously engage in these competitions. The team’s focus on documentation and platform polish indicates they are preparing for that influx. With mainnet launch, Kinitro will move from development into a growth phase, where expanding the variety of tasks and scaling up the number of miner participants could become the next objectives.
If Kinitro achieves its roadmap milestones, it could become a self-sustaining ecosystem where researchers compete to train the smartest bots, and where each improvement is rewarded on-chain. This aligns with Bittensor’s broader vision of decentralized AI development, and Kinitro will be a pioneering example of that vision applied to the robotics domain.
Kinitro is currently in an early launch phase, with the team actively working towards a full mainnet deployment of its competition system. As of late August 2025, the developers indicated they are “onboarding miners very soon” and that the validator code (for running tasks) is in the process of coming online. In fact, a project update hinted that they were targeting early September 2025 for mainnet readiness, meaning the first live competitions and reward distributions on the subnet are imminent.
Key items on Kinitro’s roadmap include:
Mainnet Competition Launch: Finalizing the implementation of the competition marketplace and evaluation system. A GitHub issue titled “Competitions system – Eval Marketplace” was opened in late August 2025, highlighting this as a feature in progress. This involves enabling validators to create tasks on-chain and a mechanism for miners to upload their models, with all participants able to view ongoing contests and their outcomes. The inclusion of a robust scoring system for evaluating agent performance is part of this launch milestone (an issue “Incorporate scoring” is being tracked as well).
Security and Reliability Enhancements: Before (and continuing after) launch, the team is hardening the platform. For example, they plan to improve logging for better debug and audit capabilities and to integrate telemetry/monitoring (using tools like PostHog) to get real-time alerts on the system’s health. Additionally, ensuring secure communications between components is on the roadmap – an issue to “harden security of backend-validator communications” is open, indicating that protecting data integrity during agent evaluation is a priority (this prevents miners from cheating or falsifying results, for instance).
Refactoring and Optimization: The team is iterating on the architecture for efficiency. The shift from a parent-child validator model to a refined platform/validator separation is one such refactor underway. These changes aim to optimize how tasks are handled and scaled – for example, to allow multiple validators to jointly handle a large competition or to make it easier to add new validators without downtime. Optimizations may also target the performance of running AI evaluations (since robotic simulations can be compute-intensive, any improvements in how the network distributes this load will be beneficial).
Documentation and Miner Tooling: Recognizing the need to attract participants, the roadmap includes creating comprehensive documentation and tutorials. There are tasks open to “add docs to [the] website” and to provide initial documentation for users. This will likely cover how new miners can create an agent (using the provided template) and join competitions, how validators can set up a node and post challenges, and general guidelines for the formats/protocols used. Improved tooling (perhaps command-line interfaces or dashboards for monitoring competitions) may also be part of the user-facing roadmap as Kinitro matures.
Future Competitions and Benchmarks: While not explicitly detailed in the issues, one can infer that once the platform is stable, a pipeline of challenge tasks (benchmarks) will be rolled out. These could range from simple tasks (like maneuvering a virtual robot through a maze) to more complex ones (multi-agent coordination, or real-world data tasks) as the community grows. The roadmap likely envisions regular competition rounds that will continuously push the state-of-the-art of the agents on the subnet. Each competition’s learnings might feed into the design of the next, creating an evolving benchmark suite (as hinted by the “continuous improvement” philosophy).
In summary, the near-term roadmap is about getting Kinitro to a fully functional state on mainnet: launching the first incentivized robotic AI challenges, ensuring the system is secure and stable, and bringing in participants. Beyond that, the project’s success will be measured by its ability to attract AI talent to continuously engage in these competitions. The team’s focus on documentation and platform polish indicates they are preparing for that influx. With mainnet launch, Kinitro will move from development into a growth phase, where expanding the variety of tasks and scaling up the number of miner participants could become the next objectives.
If Kinitro achieves its roadmap milestones, it could become a self-sustaining ecosystem where researchers compete to train the smartest bots, and where each improvement is rewarded on-chain. This aligns with Bittensor’s broader vision of decentralized AI development, and Kinitro will be a pioneering example of that vision applied to the robotics domain.
we can create environments for you (proprietary, open source, no matter), we can help you write software or train models too. anything RL-related we can help with🫡
91% now, and 39% on our harder multi task competition (MT50) that we launched just 2 days ago. Every day I wake up and see line go up. This is the power of incentives.
Kinitro is an engine for producing embodied intelligence across domains.
Robotics, humanoids, drones, materials synthesis - any environment with a clear, verifiable outcome.
You define the task and evaluation. We direct agents compete to solve it.
📩 [email protected] | Learn
we can create environments for you (proprietary, open source, no matter), we can help you write software or train models too. anything RL-related we can help with🫡
80% now.
A participant has managed to achieve a 71% success rate on our multi-task robot arm competition (Metaworld MT10).
