SAN FRANCISCO, CA, Aug 25, 2026 – Foxglove announced an AI agent layer and a new Semantic Search feature powered by NVIDIA Cosmos world models at Actuate 26. The tools let developers search, visualize, compare, debug, and curate multimodal datasets, or locate specific events in large volumes of unlabeled robot data, by describing what they need in natural language. The agent capabilities run through the Agent Sidebar, built into the Foxglove app, and the Foxglove MCP server, which connects external AI agents to the platform.
The new agent capabilities automate work that previously required manual review, turning raw multimodal data into datasets and evaluations engineers use to improve robot models.
“Robots are increasingly powered by AI, but the workflows used to develop and improve them remain largely manual,” said Adrian Macneil, co-founder and CEO of Foxglove. “Foxglove brings intelligence into the development loop, allowing engineers to describe the behavior they need to investigate or the dataset they want to create, and let agents do the work.”
The release includes four core capabilities: Agent Sidebar, Semantic Search, Comparison Mode, and Remote Access.
- Agent Sidebar runs inside the Foxglove app and coordinates data discovery, visualization, failure investigation, run comparison, and dataset curation as a single workflow.
- Semantic Search, powered by NVIDIA Cosmos, retrieves matching multimodal segments from large volumes of unlabeled data based on a natural-language description of a behavior or scenario. Each result stays linked to the synchronized sensor data, telemetry, logs, and system signals recorded around that moment.
- Comparison Mode synchronizes two or more runs so teams can see where behavior diverged across models, software versions, robots, or scenarios.
- Remote Access streams every topic from a remote robot to the browser at low latency, including camera, lidar, telemetry, and logs, letting engineers observe and debug a deployed system without traveling onsite.
NVIDIA Cosmos is a video-text embedding model trained on robotics, driving, and ego-centric human action data. General-purpose image-text models lose accuracy on complex manipulation tasks, driving scenes, and temporal action data. Foxglove and NVIDIA built the indexing and inference infrastructure for Semantic Search as a managed service, so customers do not need to run their own infrastructure to use it.
Semantic Search also works with Foxglove’s Bring-Your-Own-Storage (BYOS) architecture, indexing and searching raw logs stored in a customer’s cloud storage without duplicating the data.
“Physical AI is moving beyond recognizing objects in a frame to understanding actions as they unfold in real time,” said Amit Goel, head of robotics partner ecosystem at NVIDIA. “With Semantic Search powered by NVIDIA Cosmos world models, Foxglove is helping robotics developers search video and multimodal robot data in natural language and turn real-world context into faster debugging, evaluation and model improvement.”
Foxglove said the new capabilities depend on time-synchronized data from cameras, lidar, radar, transforms, telemetry, and logs, combined into a single record of what a robot actually did during operation.
Source: Foxglove
About Foxglove
Foxglove develops software for collecting, visualizing, debugging and managing robotics data. Founded in 2021, the company is based in San Francisco, CA. Its platform records MCAP data and imports MCAP, ROS bag, ULog and other recordings. Foxlet software uploads robot recordings and indexes them by device, time and topic. Engineers can annotate, search, share and replay recordings in configurable layouts. The software also supports live robot connections, remote visualization and teleoperation. Customers work in autonomous vehicles, defense, logistics, manufacturing, agriculture, construction, marine and health. Named users include Anduril, Shield AI, Dexterity, Wayve, Waabi, Gatik and NVIDIA. Foxglove offers cloud-hosted, customer-managed storage and on-premises deployments. It provides SDKs, APIs, webhooks, custom panels, scripts and converters. Foxglove is SOC 2 Type II compliant and supports GDPR procurement.
About Actuate 26
Actuate 26 was a two-day robotics developer conference organized by Foxglove. It was held Aug. 18-19, 2026, at Fort Mason in San Francisco. Actuate began in 2024. Program formats included keynotes, technical presentations, panels, exhibits and networking sessions. Full Access and Expo Only passes provided different levels of access to main-stage programming. Session topics included autonomous vehicles, drones, logistics, construction, biomanufacturing, delivery robots and physical AI. Speakers represented companies and research organizations, including Foxglove, NVIDIA, Wayve, Shield AI and Zipline. The event targeted robotics engineers, software developers, machine learning practitioners, founders, technical leaders and investors.
