Vortex provides an integrated platform that manages the full machine‑learning lifecycle for robotics, from versioned sensor data ingestion and multi‑framework training to simulation validation and fleet‑wide deployment. It supports PyTorch, TensorFlow, CasADi, and LangChain, exports models to ONNX or TensorRT, and automates rollout and rollback on ROS 2, Docker, Orin, and A100 hardware, with built‑in CLIP‑based dataset search for efficient curation.
Funding
Funding not disclosed
Founders
Product
Problem
Robotics teams must manage complex machine‑learning pipelines that span raw sensor ingestion, multi‑framework training, simulation validation, and fleet‑wide deployment, often using disparate tools that hinder reproducibility and scalability.
Solution
Vortex offers an integrated platform that orchestrates the entire ML model lifecycle for real‑world robots. It ingests and version‑controls sensor data from any source, then launches training jobs across PyTorch, TensorFlow, CasADi, and LangChain with configurable hyperparameters and live log streaming. Models can be validated in simulators such as NVIDIA Isaac Sim, Gazebo, or Isaac Gym before being exported to ONNX, TensorRT, or native formats. The platform packages the optimized artifacts for deployment to ROS 2, Docker, Orin, or A100 hardware across an entire robot fleet, supporting rollbacks when needed. Built‑in visual semantic search powered by CLIP and LLM query decomposition helps users curate datasets efficiently.
Target Audience
Primary customers are robotics companies and research labs that develop autonomous perception, planning, and control systems and need a unified solution to train, validate, and deploy ML models at scale across robot fleets.
Features
- Versioned, deduplicated data ingestion pipeline that connects to arbitrary sensor sources
- Multi‑framework training orchestration with live monitoring for PyTorch, TensorFlow, CasADi, and LangChain
- Simulator integration for policy validation in NVIDIA Isaac Sim, Gazebo, and Isaac Gym
- Export utilities supporting ONNX, TensorRT, and native weight formats for diverse hardware targets
- Fleet deployment tools with ROS 2, Docker, Orin, and A100 compatibility, including automated rollback
- CLIP‑based visual semantic search combined with LLM query decomposition for agentic dataset curation
- End‑to‑end perception stack (DROID‑SLAM, Mask R‑CNN) and policy/control modules (Isaac Gym, RL/IL) within the same platform