Positronic Robotics trains control models for cleaning robots on Nebius Compute Cloud. Its customer story describes several virtual machines with H100 GPUs and containerized training using distributed parallel computing. One server runs each 12-to-16-hour training cycle. Scripts automate tasks and shut down the server after training. Nebius also provides storage and management for around one terabyte of training data. Data gathered through VR tools and sensors is processed and prepared for the models. Human demonstrations teach robots tasks such as wiping surfaces. Nebius reports that the trained models support an end-to-end autonomous cleaning system.
The team plans to expand data collection by equipping cleaning staff with specialized tools. It is exploring simulators on Nebius servers for model validation and larger-scale parallel training. This remains prospective use described in the customer story. In April 2026, Nebius again named Positronic as running on its infrastructure. Positronic launched the Physical AI Leaderboard with Nebius and Toloka as founding consortium partners, measuring real-robot picking throughput and time between failures. Nebius announced that it would provide infrastructure for robot-model fine-tuning and evaluation. A joint webinar advertised demonstrations of containerized GPU jobs on Nebius AI Cloud for simulation, synthetic dataset generation and model evaluation.
The recorded workshop describes Positronic's data conversion and training running as serverless jobs, with policy inference through an endpoint. A live demonstration connected a robot in Positronic's lab to a policy served on Nebius. The speaker reported roughly 300 milliseconds of round-trip latency, including inference, for action chunks covering 1.5 to 2 seconds. A complete training run was not performed during the webinar. Separately, a repository issue reports successful RoboLab/Isaac Sim rendering and a full evaluation replay on a Nebius L40S Compute VM. Serverless evaluation tests failed because graphics driver components were missing. The issue describes provisioned VMs as the working evaluation path while the serverless blocker remains unresolved. The workflow README records successful image pulls, L40S scheduling and cache reuse on Nebius serverless jobs. It also documents the graphics-library setup needed for VM evaluation.
positronic/workflows/nebius/README.md at main · Positronic-Robotics/positronic · GitHub
18 Oct 2026 · github.com · Web page
Positronic Robotics documents CPU dataset conversion, H100 training jobs and serverless policy endpoints in its own Nebius project. Training inputs stream from object storage. Endpoint tests report an approximately 90-second idle connection closure, addressed with WebSocket pings and gRPC keepalive. The README also reports approximately 6 ms gRPC and 60 ms WebSocket round trips for an 846 KiB observation through Nebius’s managed front.
positronic/workflows/nebius/README.md at main · Positronic-Robotics/positronic · GitHub18 Oct 2026 · github.com · Web pagePositronic Robotics documents CPU dataset conversion, H100 training jobs and serverless policy endpoints in its own Nebius project. Training inputs stream from object storage. Endpoint tests report an approximately 90-second idle connection closure, addressed with WebSocket pings and gRPC keepalive. The README also reports approximately 6 ms gRPC and 60 ms WebSocket round trips for an 846 KiB observation through Nebius’s managed front.
positronic/workflows/nebius at main · Positronic-Robotics/positronic · GitHub
18 Oct 2026 · github.com · Web page
Positronic Robotics documents a Nebius H100 training workflow supporting ACT, SmolVLA, OpenPI and GR00T. Input datasets stream through a read-only Mountpoint-S3 mount. Checkpoint outputs use pos3 because vendor checkpoint savers can require symlinks that Mountpoint-S3 does not support.
positronic/workflows/nebius at main · Positronic-Robotics/positronic · GitHub18 Oct 2026 · github.com · Web pagePositronic Robotics documents a Nebius H100 training workflow supporting ACT, SmolVLA, OpenPI and GR00T. Input datasets stream through a read-only Mountpoint-S3 mount. Checkpoint outputs use pos3 because vendor checkpoint savers can require symlinks that Mountpoint-S3 does not support.
RoboLab eval on Nebius Serverless: jobs get compute-only driver injection, Isaac Sim cannot create a Vulkan device · Issue #484 · Positronic-Robotics/positronic · GitHub
18 Oct 2026 · github.com · Web page
Repository contributor vertix reports successful RoboLab/Isaac Sim rendering and a full evaluation replay on a Nebius L40S Compute VM. Nebius Serverless AI evaluation jobs fail at startup because graphics driver components are missing. Provisioned VMs remain the working evaluation path.
RoboLab eval on Nebius Serverless: jobs get compute-only driver injection, Isaac Sim cannot create a Vulkan device · Issue #484 · Positronic-Robotics/positronic · GitHub18 Oct 2026 · github.com · Web pageRepository contributor vertix reports successful RoboLab/Isaac Sim rendering and a full evaluation replay on a Nebius L40S Compute VM. Nebius Serverless AI evaluation jobs fail at startup because graphics driver components are missing. Provisioned VMs remain the working evaluation path.
Nebius Customer story
10 Mar 2025 · Nebius · Web page
Nebius describes Positronic Robotics training robot-control models on H100 virtual machines. Its control system uses an Action Chunking Transformer trained as a conditional variational autoencoder to capture variation in human demonstrations. Cartesian movement representations allow policies and demonstrations to work across different robotic arms.
Nebius Customer story10 Mar 2025 · Nebius · Web pageNebius describes Positronic Robotics training robot-control models on H100 virtual machines. Its control system uses an Action Chunking Transformer trained as a conditional variational autoencoder to capture variation in human demonstrations. Cartesian movement representations allow policies and demonstrations to work across different robotic arms.
Fast experiment loops for robotics: simulation, evaluation, and dataset generation
9 Oct 2026 · nebius.com · Web page
Nebius's recorded workshop describes Positronic's data conversion and training running as serverless jobs and policy inference through an endpoint. A live demonstration connects a robot in Positronic's lab to a policy served on Nebius. The speaker reports roughly 300 milliseconds of round-trip latency, including inference, with action chunks covering 1.5 to 2 seconds. The workshop uses prepared training examples without performing a complete training run.
Fast experiment loops for robotics: simulation, evaluation, and dataset generation9 Oct 2026 · nebius.com · Web pageNebius's recorded workshop describes Positronic's data conversion and training running as serverless jobs and policy inference through an endpoint. A live demonstration connects a robot in Positronic's lab to a policy served on Nebius. The speaker reports roughly 300 milliseconds of round-trip latency, including inference, with action chunks covering 1.5 to 2 seconds. The workshop uses prepared training examples without performing a complete training run.
Positronic Robotics provides independent evaluation of robot AI models for research labs and robotics developers. Teams submit model weights or a hosted endpoint, and Positronic runs comparisons on its robots under matched conditions, returning videos, scores and results the same day. Its services help customers such as Runway assess model checkpoints during training without managing evaluation hardware and operators themselves. The company also runs PhAIL, a public Physical AI Leaderboard, and publishes an open-source Python toolkit for building machine-learning robotics systems.
Positronic Robotics trains control models for cleaning robots on Nebius Compute Cloud. Its customer story describes several virtual machines with H100 GPUs and containerized training using distributed parallel computing. One server runs each 12-to-16-hour training cycle. Scripts automate tasks and shut down the server after training. Nebius also provides storage and management for around one terabyte of training data. Data gathered through VR tools and sensors is processed and prepared for the models. Human demonstrations teach robots tasks such as wiping surfaces. Nebius reports that the trained models support an end-to-end autonomous cleaning system.
The team plans to expand data collection by equipping cleaning staff with specialized tools. It is exploring simulators on Nebius servers for model validation and larger-scale parallel training. This remains prospective use described in the customer story. In April 2026, Nebius again named Positronic as running on its infrastructure. Positronic launched the Physical AI Leaderboard with Nebius and Toloka as founding consortium partners, measuring real-robot picking throughput and time between failures. Nebius announced that it would provide infrastructure for robot-model fine-tuning and evaluation. A joint webinar advertised demonstrations of containerized GPU jobs on Nebius AI Cloud for simulation, synthetic dataset generation and model evaluation.
The recorded workshop describes Positronic's data conversion and training running as serverless jobs, with policy inference through an endpoint. A live demonstration connected a robot in Positronic's lab to a policy served on Nebius. The speaker reported roughly 300 milliseconds of round-trip latency, including inference, for action chunks covering 1.5 to 2 seconds. A complete training run was not performed during the webinar. Separately, a repository issue reports successful RoboLab/Isaac Sim rendering and a full evaluation replay on a Nebius L40S Compute VM. Serverless evaluation tests failed because graphics driver components were missing. The issue describes provisioned VMs as the working evaluation path while the serverless blocker remains unresolved. The workflow README records successful image pulls, L40S scheduling and cache reuse on Nebius serverless jobs. It also documents the graphics-library setup needed for VM evaluation.
positronic/workflows/nebius/README.md at main · Positronic-Robotics/positronic · GitHub
18 Oct 2026 · github.com · Web page
Positronic Robotics documents CPU dataset conversion, H100 training jobs and serverless policy endpoints in its own Nebius project. Training inputs stream from object storage. Endpoint tests report an approximately 90-second idle connection closure, addressed with WebSocket pings and gRPC keepalive. The README also reports approximately 6 ms gRPC and 60 ms WebSocket round trips for an 846 KiB observation through Nebius’s managed front.
positronic/workflows/nebius/README.md at main · Positronic-Robotics/positronic · GitHub18 Oct 2026 · github.com · Web pagePositronic Robotics documents CPU dataset conversion, H100 training jobs and serverless policy endpoints in its own Nebius project. Training inputs stream from object storage. Endpoint tests report an approximately 90-second idle connection closure, addressed with WebSocket pings and gRPC keepalive. The README also reports approximately 6 ms gRPC and 60 ms WebSocket round trips for an 846 KiB observation through Nebius’s managed front.
positronic/workflows/nebius at main · Positronic-Robotics/positronic · GitHub
18 Oct 2026 · github.com · Web page
Positronic Robotics documents a Nebius H100 training workflow supporting ACT, SmolVLA, OpenPI and GR00T. Input datasets stream through a read-only Mountpoint-S3 mount. Checkpoint outputs use pos3 because vendor checkpoint savers can require symlinks that Mountpoint-S3 does not support.
positronic/workflows/nebius at main · Positronic-Robotics/positronic · GitHub18 Oct 2026 · github.com · Web pagePositronic Robotics documents a Nebius H100 training workflow supporting ACT, SmolVLA, OpenPI and GR00T. Input datasets stream through a read-only Mountpoint-S3 mount. Checkpoint outputs use pos3 because vendor checkpoint savers can require symlinks that Mountpoint-S3 does not support.
RoboLab eval on Nebius Serverless: jobs get compute-only driver injection, Isaac Sim cannot create a Vulkan device · Issue #484 · Positronic-Robotics/positronic · GitHub
18 Oct 2026 · github.com · Web page
Repository contributor vertix reports successful RoboLab/Isaac Sim rendering and a full evaluation replay on a Nebius L40S Compute VM. Nebius Serverless AI evaluation jobs fail at startup because graphics driver components are missing. Provisioned VMs remain the working evaluation path.
RoboLab eval on Nebius Serverless: jobs get compute-only driver injection, Isaac Sim cannot create a Vulkan device · Issue #484 · Positronic-Robotics/positronic · GitHub18 Oct 2026 · github.com · Web pageRepository contributor vertix reports successful RoboLab/Isaac Sim rendering and a full evaluation replay on a Nebius L40S Compute VM. Nebius Serverless AI evaluation jobs fail at startup because graphics driver components are missing. Provisioned VMs remain the working evaluation path.
Nebius Customer story
10 Mar 2025 · Nebius · Web page
Nebius describes Positronic Robotics training robot-control models on H100 virtual machines. Its control system uses an Action Chunking Transformer trained as a conditional variational autoencoder to capture variation in human demonstrations. Cartesian movement representations allow policies and demonstrations to work across different robotic arms.
Nebius Customer story10 Mar 2025 · Nebius · Web pageNebius describes Positronic Robotics training robot-control models on H100 virtual machines. Its control system uses an Action Chunking Transformer trained as a conditional variational autoencoder to capture variation in human demonstrations. Cartesian movement representations allow policies and demonstrations to work across different robotic arms.
Fast experiment loops for robotics: simulation, evaluation, and dataset generation
9 Oct 2026 · nebius.com · Web page
Nebius's recorded workshop describes Positronic's data conversion and training running as serverless jobs and policy inference through an endpoint. A live demonstration connects a robot in Positronic's lab to a policy served on Nebius. The speaker reports roughly 300 milliseconds of round-trip latency, including inference, with action chunks covering 1.5 to 2 seconds. The workshop uses prepared training examples without performing a complete training run.
Fast experiment loops for robotics: simulation, evaluation, and dataset generation9 Oct 2026 · nebius.com · Web pageNebius's recorded workshop describes Positronic's data conversion and training running as serverless jobs and policy inference through an endpoint. A live demonstration connects a robot in Positronic's lab to a policy served on Nebius. The speaker reports roughly 300 milliseconds of round-trip latency, including inference, with action chunks covering 1.5 to 2 seconds. The workshop uses prepared training examples without performing a complete training run.
Positronic Robotics provides independent evaluation of robot AI models for research labs and robotics developers. Teams submit model weights or a hosted endpoint, and Positronic runs comparisons on its robots under matched conditions, returning videos, scores and results the same day. Its services help customers such as Runway assess model checkpoints during training without managing evaluation hardware and operators themselves. The company also runs PhAIL, a public Physical AI Leaderboard, and publishes an open-source Python toolkit for building machine-learning robotics systems.
