Bittensor subnet news

Nepher Robotics@nepher_robotics

Humanoid Run Jump Phase 3 is officially in the books, with a winning score of 0.8245. Across three phases, miners have continued to push the same Unitree G1 humanoid to perform better on the same task: • Phase 1: 0.6750 • Phase 2: 0.7735 • Phase 3: 0.8245 Every phase used the same hidden benchmark tracks and open evaluation harness, with independent validators scoring each policy inside isolated sandboxes. And the competition is getting tighter. Just 0.0018 separated first and second place in Phase 3. Now, the Unitree Go2 LiDAR Maze Challenge is running on the same evaluation pipeline. This time, miners are training a high-level planner on top of a frozen locomotion policy, competing to navigate deterministic benchmark mazes. Submissions close 18 October. 🔗: nepher.ai/tournament Same evaluation standards, tougher competition, and better-performing robot policies with every phase.

Full post on TaosisOpen on X ↗

Nepher Robotics@nepher_robotics

Nepher Robotics looking for robot manufacturer partners! Please stay in touch.

Full post on TaosisOpen on X ↗

Nepher Robotics@nepher_robotics

Excited to join the European robotics community at @NVIDIAGTC Berlin next week (Oct 20–22)! We're Nepher Robotics — the platform that turns NVIDIA Omniverse + Isaac Lab into production-ready robot policies, from simulation to real hardware in days, not months. Open tournaments. Standardized EnvHub environments. Independent evaluation. Sim-to-real packages. If you're building or deploying robots and want to talk partnerships, benchmarks, or talent — let's connect in Berlin or online. nepher.ai #NVIDIAGTC #PhysicalAI #IsaacLab #Robotics

Full post on TaosisOpen on X ↗

Nepher Robotics@nepher_robotics

🚀 Training & evaluating VLA (Vision-Language-Action) and WAM (World Action Models) just got a major upgrade with @nvidiaomniverse’s Isaac Sim + Isaac Lab. While MuJoCo remains solid for contact-rich physics and quick academic baselines, Isaac Sim delivers clear edges for modern foundation models: ✅ Photorealistic RTX rendering → far better visual sim-to-real transfer (critical for vision-heavy VLAs & WAMs) ✅ Massive GPU parallelism (thousands of envs) → dramatically faster policy training & large-scale evaluation ✅ Built-in synthetic data + domain randomization via Replicator → richer, more diverse training signals ✅ Seamless integration with NVIDIA’s Physical AI stack (Cosmos, GR00T, Isaac ROS) for end-to-end workflows WAMs especially benefit — their world-dynamics priors shine when the simulator can deliver realistic multi-modal observations at scale. MuJoCo is still great for pure dynamics research. But when you’re building generalist robot policies that must see, reason, and act in the real world… Isaac Sim is the platform. #Robotics #PhysicalAI #VLA #WAM #IsaacSim

Full post on TaosisOpen on X ↗

Nepher Robotics@nepher_robotics

A new robot just joined SN49: the Unitree Go2. The Unitree Go2 LiDAR Maze Navigation tournament is now live. Miners train a policy that guides the Go2 through enclosed mazes using nothing but a 360° LiDAR scan. There are no cameras and no map, so the robot has to read the walls around it, decide where to go next and reach the goal without falling. If that sounds familiar, it should. Our Leatherback LiDAR Real Maze tournament drew 337 submissions from 97 miners, and the winning policy scored 0.9280. This time the same challenge moves from a wheeled car to a four-legged robot, where every turn also has to keep a walking body stable. Miners train the high-level planner that chooses the path, on top of a frozen walking policy, and they get everything needed to start: • An Isaac Lab training environment and the Unitree Go2 model • Training and evaluation scripts • The frozen locomotion policy • Benchmark mazes for reproducible scoring Final scores come from deterministic benchmark mazes, so every policy is judged on exactly the same courses by independent validators. The top three share a reward pool of $13.9k under our 70/20/10 split, and submissions close OCT 18 🔗: nepher.ai/tournament/tournaments/997ea97f-4371-…

Full post on TaosisOpen on X ↗

Nepher Robotics@nepher_robotics

In July, a circle that looked perfect in simulation fell apart the moment it ran on our real SO-ARM101. Today that same arm writes NEPHER and traces our logo, first in simulation and then on real hardware, following the same path stroke for stroke. It's a big step up from the $TAO symbol we shared in August. A full word and a logo mean many separate strokes, sharp corners and smooth curves, and the arm has to finish each one cleanly before moving precisely to the next. Small errors that would go unnoticed in a single shape add up fast across six letters. This is the gap every robotics team fights to close, getting motion that works in simulation to hold up on a real machine. It's also the loop Nepher is building into SN49, where robot policies developed in simulation earn their way onto real hardware. From a failed circle to our own name in a little over two months. More from the real arm soon.

Full post on TaosisOpen on X ↗

Nepher Robotics@nepher_robotics

Humanoid Run Jump Phase 3 came down to a margin of 0.0018. That's how close first and second place finished in the first tournament under our new 70/20/10 reward split. Under winner-takes-all, second place would have walked away with nothing. This time, the top three all get paid. • 1st: 5G1VVfTHDEuW4j7x87pjpNUkKX7CM2eH3u8qUzCPQxvTrwGW with a score of 0.8245, earning 4130 Alpha • 2nd: 5GHRBj5g8BkmuBv2ignmk4VtjxdHjVxLBGQUf69DfnT4QREH with a score of 0.8227, earning 1180 Alpha • 3rd: 5D9vwJyWrdsgEtqb64L3ZyL3RxzKfSu2opTHzfQyzp3EnX1N with a score of 0.8036, earning 590 Alpha Every miner trained a Unitree G1 humanoid to run, jump over obstacles and stay stable, and every policy was scored on hidden benchmark tracks by independent validators. The winning score also keeps climbing from one phase to the next: 0.6750 in Phase 1, 0.7735 in Phase 2 and 0.8245 in Phase 3. Congratulations to the winners, and thank you to every miner who competed. 🔗: nepher.ai/tournament/tournaments/373d2333-fe83-…

Full post on TaosisOpen on X ↗

Nepher Robotics@nepher_robotics

Last month, Nepher Verified became a selectable certification for robot manufacturers, listed alongside CE, UL, ISO, TÜV, FCC and FDA. Our CEO and co-founder, Mahdi Eghbali (@eghbali_mahdii), PhD, explains what that means for robotics. When a robot maker says their hardware or policy performs, buyers usually have to take their word for it. Nepher Verified replaces that self-reported claim with an independent score, produced by the same evaluation pipeline that judges every tournament on SN49. Every certificate comes with: • A public verification ID • A public registry entry • An official document anyone can check in seconds It's already live in the US, and expansion into Europe and Asia is underway. The same infrastructure that scores miners' robot policies is now becoming a trust layer for the robotics industry. Manufacturers can apply for certification or sponsor a tournament here: nepher.ai/nepher-verified

Full post on TaosisOpen on X ↗