OpenRoboto

Scroll sideways for more →

OpenRoboto is a robotics model competition that runs as subnet 80 on the Bittensor network. Miners fine-tune an open vision-language-action base model, currently π0.5 from Physical Intelligence, publish the resulting weights on Hugging Face, and commit them on chain; the subnet scores every submission on the LIBERO manipulation benchmark in simulation and pays the top performers through Bittensor token emissions ^[1][3]. The project announced its official mainnet launch on August 6, 2026, describing its goal as "open weights, open data and open benchmarks, coordinated through Bittensor to continuously improve robotics models" ^[1]. As of August 7, 2026 it is an early-stage system: one competition round was live, two models had cleared the ranking bar, and the promised open data pool was still empty ^[7][12].

Background

Most capable robot manipulation policies today are trained inside a small number of private labs, and OpenRoboto presents itself as an open counterweight: every submitted model is a downloadable artifact under open weights, every score is published with the inputs needed to recompute it, and the best model of each round becomes the shared base for the next ^[2][3]. The design borrows the incentive structure of Bittensor, a blockchain network in which specialized "subnets" pay contributors in the network's TAO-linked tokens for measurable machine-learning work. Under Bittensor's dynamic TAO (dTAO) system, each subnet also has its own tradable "alpha" token, priced through an automated pool against TAO ^[19].

The public record of the project is recent. The openroboto-ai GitHub organization was created on June 29, 2026 and lists its location as China; its Hugging Face organization and empty "datapool" storage bucket followed in late July ^[6][11][12]. The subnet contact published on the block explorer taostats is riccardo@openroboto.ai ^[8]. No named founders, corporate entity, or funding announcements were identified as of August 7, 2026, and no coverage in major AI or crypto outlets was identified as of the same date.

How the subnet works

OpenRoboto is structured as a continuous open competition rather than a task marketplace. Miners pull the current base model, train it with any recipe they like (LoRA included), and must merge the result into a complete checkpoint before submitting; bare adapters are rejected by a CPU pre-check before any GPU time is spent ^[3][4]. A submission is a single on-chain commitment of at most 512 bytes that binds together the miner's hotkey, the exact Hugging Face commit hash of the uploaded weights, the round number, and the transaction hash of the evaluation fee ^[3].

The fee is paid by burning TAO (0.1 TAO per submission as of August 7, 2026, with the rate published in a public control.json file). Burning rather than transferring means no one receives the fee, which the project argues removes any incentive for the operator to farm submissions; the burn and the commitment must land within 10 blocks (about 2 minutes) of each other as an anti-replay rule, and a rejected submission's fee is not refunded ^[3]. Two further rules limit abuse: each hotkey gets one commitment per round, and the Hugging Face repository name must end with the last 12 characters of the submitting hotkey's address so that impersonation is detectable ^[3].

Evaluation seeds are derived from two public randomness sources that neither the miner nor the operator controls: the Bittensor block hash of the block containing the commitment, and the drand distributed randomness beacon ^[3][18]. The backend polls the chain every 60 seconds, verifies payment, and queues the model for a benchmark worker that loads the pinned Hugging Face revision and runs the LIBERO suites in MuJoCo ^[3].

Ranking uses a king-of-the-hill challenge system. The first scored miner becomes champion; each later submission must beat the champion's average score by a published margin (0.02 by default) or it does not appear on the board at all. A copy of the champion's weights ties and therefore loses, which blocks weight-copying attacks. The board is capped at three entries, and a lightweight validator translates it into on-chain emission weights of 70, 20, and 10 percent for ranks one through three ^[3]. Anyone can confirm the emitted weights with standard Bittensor tooling (btcli subnet metagraph, netuid 80) ^[3].

The project is unusually direct about its trust limitations. Weight fingerprinting (sha256 hashes of every uploaded shard, used to catch resubmitted copies) runs in detection mode only, and scores are produced by the operator's own benchmark workers rather than by independent validators; what keeps the operator honest, the documentation argues, is that the harness is open source and every input is pinned, so any published score can be recomputed offline ^[3][5].

Benchmark and evaluation harness

Submissions are scored across six LIBERO suites per round: the four standard suites plus two "swap" perturbation variants ^[3].

Scroll sideways for more →

LIBERO is a 2023 lifelong robot learning benchmark built on human demonstrations ^[15]. Because those demonstrations are public, a model trained through imitation learning can score well by memorizing them: OpenRoboto says an internal red-team study found the four base suites could be saturated for roughly 50 dollars of GPU time that way, while the swap perturbations showed no transfer from the attack, which is why they are included and, in the production scoring profile, double-weighted ^[3][5]. The perturbation suites draw on LIBERO-PRO, a 2025 evaluation study of memorization in vision-language-action models ^[16]; OpenRoboto's published measurements on the official π0.5 checkpoint show that rephrasing the instruction costs the model almost nothing (97.5 percent success) while swapping object positions collapses it to 25.8 percent ^[5]. The harness also supports the larger LIBERO-plus robustness benchmark ^[5][17].

The open-source evaluation harness fans tasks out across eight GPUs and follows openpi's official LIBERO protocol. The project reports reproducing 97.2 percent average success on the reference pi05_libero checkpoint against the officially published 96.85 percent, and publishes the alignment data ^[5]. Per the published benchmark spec, each suite runs with 50 trials per task at 20 Hz control (7-DoF joint deltas plus gripper) with wrist and front RGB observations ^[2].

Base model

The starting point for every miner is π0.5, an open-weights robot foundation model in the π0 family released by Physical Intelligence through its openpi repository ^[13][14]. OpenRoboto's documentation puts the model at roughly 3 billion parameters and accepts submissions in openpi's JAX or PyTorch checkpoint formats, with an architecture check that requires a parameter count between 2.5 and 4.5 billion ^[3]. The subnet's stated long-term mechanism is compounding: when a challenger wins a round, its weights become the next round's base, so improvements accumulate in public rather than in any one lab ^[2][3].

Launch and early activity

Per the project's own roadmap, a testnet phase in the third quarter of 2026 shipped a minimal π0.5-plus-LIBERO loop, a live leaderboard, and a first champion ^[2]. The mainnet launch was announced on X on August 6, 2026 ^[1]. On August 7, the public API showed Round 01 live on mainnet with 15 total submissions, of which two had cleared the challenge bar: the champion, submitted by miner UID 39, scored 0.749 average success against the π0.5 baseline's 0.503 on the same seeded evaluation, and the runner-up scored 0.706 ^[7]. Both entries followed the required naming convention, with full checkpoints public on Hugging Face ^[7]. Taostats showed 256 registered keys and 10 active validators on the subnet the same day, and listed the subnet's alpha token at a market value equivalent to roughly 18,000 to 19,000 TAO, a few million US dollars at that day's TAO price; the figure swings substantially with the token's intraday trading ^[8].

The roadmap's later stages remain unbuilt: a "real-robot gate" in which policies must run on a physical robot is slated for the first quarter of 2027, a self-built benchmark to replace LIBERO for the second quarter, and factory-driven policy tournaments for the third ^[2]. As of August 7, 2026, nothing in the live system involved physical robots, and the scoring loop ran entirely in simulation.

History of netuid 80

Netuid 80 predates OpenRoboto and has carried two other identities. A May 2025 snapshot of taostats records netuid 80 as "AI Factory," a self-described specialized AI development subnet ^[9]. The current registration record dates to December 22, 2025 (block 7,151,800, at a registration cost of about 220 TAO), and a January 2026 snapshot shows that registration operating as "Dogelayer," a merged-mining project ^[8][10]. By August 2026 the same registration carried the OpenRoboto identity, with a different owner coldkey than in January ^[8][10]. Whether the OpenRoboto team acquired the slot from Dogelayer's operator or the subnet was transferred some other way is not publicly documented.

Openness in practice

The competition code is public: the miner toolkit, protocol helpers, weight-setting validator, and the full evaluation harness are published under MIT licenses, and the two model repositories on the project's Hugging Face organization carry Apache 2.0 licenses ^{[4][5][6][11]}. Other parts of the "open weights, open data, open benchmarks" vision are still aspirational. The "Open Data Pool" advertised on the website, a Hugging Face storage bucket meant to hold curated LIBERO demonstrations and crowd-collected simulation episodes, contained zero files as of August 7, 2026, with the site listing dataset sizes and licenses as "TBA" ^[2][12]. Held-out task data and the scoring service deployment are explicitly outside the public repositories ^[4]. The project also publishes no information about who runs it beyond the contact email and the GitHub organization's stated location.

OpenRoboto is not the only attempt to attach robotics to Bittensor incentives: Nepher Robotics operates subnet 49 as a tournament platform in which miners submit policies that validators score in sandboxed Isaac Lab containers ^[20]. Both projects apply the same underlying bet, that open-source competition with token rewards can close some of the gap with private embodied AI labs, whose best manipulation and humanoid robot systems remain closed. Whether a benchmark-driven subnet can produce robot policies that matter outside simulation is the question OpenRoboto's own 2027 real-robot roadmap has yet to answer.

See also

• Bittensor
• Physical Intelligence
• π0
• LIBERO
• Robot foundation model
• Vision-language-action model
• Open weights

References

1. ^OpenRoboto (@openroboto). Mainnet launch announcement. X, August 6, 2026. x.com/...2085369343253332175
2. ^OpenRoboto. "Robot intelligence built in the open" (website, benchmark, data pool, and roadmap pages). Accessed August 7, 2026. openroboto.ai
3. ^OpenRoboto. "OpenRoboto Subnet: Protocol & Incentive Mechanism" (SUBNET_OVERVIEW.md). GitHub, accessed August 7, 2026. github.com/...SUBNET_OVERVIEW.md
4. ^OpenRoboto. "OpenRoboto Miner and Protocol" (repository README). GitHub, accessed August 7, 2026. github.com/...openroboto-subnet
5. ^OpenRoboto. "OpenRoboto Validator: Parallel LIBERO evaluation" (repository README). GitHub, accessed August 7, 2026. github.com/...openroboto-evaluation
6. ^GitHub. openroboto-ai organization page (created June 29, 2026; location and repositories). Accessed August 7, 2026. github.com/openroboto-ai
7. ^OpenRoboto public API. Round and leaderboard endpoints, accessed August 7, 2026. api.openroboto.ai/...leaderboard
8. ^Taostats. "Subnet 80: OpenRoboto" (registration record, metagraph, and token data). Accessed August 7, 2026. taostats.io/...80
9. ^Internet Archive Wayback Machine. Capture of taostats.io/subnets/80 ("AI Factory"), May 21, 2025. web.archive.org/...80
10. ^Internet Archive Wayback Machine. Capture of taostats.io/subnets/80/chart ("dogelayer"), January 22, 2026. web.archive.org/...chart
11. ^Hugging Face. openroboto-ai organization (models and licenses). Accessed August 7, 2026. huggingface.co/openroboto-ai
12. ^Hugging Face. openroboto-ai/datapool storage bucket (created July 20, 2026). Accessed August 7, 2026. huggingface.co/...datapool
13. ^Physical Intelligence. "openpi" (open-source robot learning repository with π0.5 checkpoints). GitHub, accessed August 7, 2026. github.com/...openpi
14. ^Physical Intelligence. "π0.5: a Vision-Language-Action Model with Open-World Generalization." arXiv, April 2025. arxiv.org/...2504.16054
15. ^Liu, Bo, et al. "LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning." arXiv, June 2023. arxiv.org/...2306.03310
16. ^"LIBERO-PRO: Towards Robust and Fair Evaluation of Vision-Language-Action Models Beyond Memorization." arXiv, October 2025. arxiv.org/...2510.03827
17. ^"LIBERO-Plus: In-depth Robustness Analysis of Vision-Language-Action Models." arXiv, October 2025. arxiv.org/...2510.13626
18. ^drand. "Distributed randomness beacon." Accessed August 7, 2026. drand.love
19. ^Learn Bittensor. "Subnet Pool" (Dynamic TAO concept documentation). Accessed August 7, 2026. learnbittensor.org/...subnet-pool
20. ^Nepher Robotics. "Nepher Robotics Subnet" (Bittensor subnet 49 repository README). GitHub, accessed August 7, 2026. github.com/...nepher-subnet