Terra
Product · SaaS / cloud software by Labelbox
The data engine for physical AI and robotics at scale.
Terra is a robotics data engine that provides synchronized multi-camera demonstrations and expert teleoperation data for training physical AI. It uses seven specialized hardware rigs and an AI diversity engine to deliver 1PB+ of validated training data across diverse real-world environments.
- Languages
- English
Category: Data Labeling
- Platforms
- Web-based platform for data management and validation
- Integrations
- Alignerr network for expert human data collection
- Limits
- Up to 4 synced camera streams per episode
- Data Produced
- 1PB+ (as of Jan 2026)
- Rig Configurations
- 7 setups
- Max Camera Streams
- 4 synced streams
- 1PB+ robotics data produced as of January 2026
- 7 hardware-synchronized capture rig configurations
- Up to 4 synced camera streams per episode
- Unified episode format from pre-training to evals
Overview
Terra is Labelbox's specialized data engine for robotics, designed to generate high-quality training data for physical AI at scale. It provides egocentric and teleoperated demonstrations captured on seven distinct hardware-synchronized rig configurations, ranging from head-mounted stereo cameras to instrumented grippers. The platform solves the data scarcity problem in robotics by using a global network of operators to record real-world tasks across diverse environments like kitchens, warehouses, and labs. Every episode is ingested into a unified format, validated by an AI diversity engine, and delivered with synchronized action and state streams. Terra supports the entire development lifecycle, from diverse video for pre-training to expert-graded demonstrations for imitation learning and standardized evaluation rubrics. As of January 2026, the engine has produced over 1PB of robotics training data for AI research and engineering teams.
Data Generation and Management
- Data Collection Rigs: The platform provides egocentric and teleoperated demonstrations captured on purpose-built hardware rigs for robotics training.
- Unified Episode Format: Every rig writes to the same episode format, ensuring data for pre-training, post-training, and evaluation is consistent.
- Synchronized Data Streams: The system delivers synchronized multi-camera video, action streams, and state streams in formats expected by training stacks.
- AI Diversity Engine: An AI diversity engine auto-categorizes episodes by environment, object, task, and operator to identify and fill coverage gaps.
Hardware Configurations
- Ego Stereo Rig: The Ego Stereo rig uses two head-mounted synced cameras to capture RGB and depth data for 3D scene understanding and navigation.
- Trio + Gripper Rig: The Trio + Gripper rig combines three head cameras with an instrumented gripper to provide action-labeled demonstrations.
- Ego Trio Rig: The Ego Trio rig uses three head and peripheral cameras to provide a wide field of view for full-context manipulation tasks.
- Teleoperation Rig: The Teleoperation rig records robot-mounted video along with joint states and operator actions for imitation learning.
- Overhead Rig: The Overhead rig uses one to two fixed cameras above the workspace for bird's-eye view and bimanual tracking.
Training and Evaluation Pipeline
- Pre-training Data: The platform provides diverse egocentric and third-person video with auto-labels for scene, object, and action segments.
- Post-training Data: Expert teleoperation and instrumented-gripper demonstrations include success labels for imitation learning and RL fine-tuning.
- Evaluation Infrastructure: Standardized manipulation and navigation tasks are run against policies with scoring rubrics and per-environment reports.
Use cases
- Robotics teams use Ego Stereo rigs to capture 3D scene data for training autonomous navigation models.
- Research labs use the Trio + Gripper rig to collect action-labeled demonstrations for imitation learning in manipulation tasks.
- AI engineers use the diversity engine to identify gaps in their training sets and steer collection toward underrepresented environments.
- Developers use the standardized evaluation protocol to run their policies against scoring rubrics and generate per-environment reports.
Specifications
- Deployment
- Cloud-based data engine with physical hardware rig deployment for data collection
- Platforms
- Web-based platform for data management and validation
- Integrations
- Alignerr network for expert human data collection
- APIs and SDKs
- Delivers data in formats compatible with RL training and evaluation infrastructure
- Limits
- Up to 4 synced camera streams per episode
- Data Produced
- 1PB+ (as of Jan 2026)
- Rig Configurations
- 7 setups
- Max Camera Streams
- 4 synced streams
- Camera Modalities
- RGB, RGB + depth
- Data Formats
- Synchronized video, action, and state streams
- Collection Environments
- Kitchen, office, warehouse, retail, home, lab
- Rig Types
- Ego mono, Ego stereo, Ego trio, Ego + wrist, Trio + gripper, Overhead, Teleoperation
How it works
- Ingest Raw Streams: Raw data streams land from operators worldwide using purpose-built, hardware-synchronized rigs.
- Categorize and Tag: The system auto-tags episodes by environment, objects, task, and operator style.
- Validate Quality: Automated checks and human review on edge cases ensure every episode meets quality standards before shipping.
- Steer Collection: The diversity engine identifies gaps in coverage and redirects collection efforts to fill those specific needs.
FAQ
Does Labelbox build robots?
No, Labelbox does not build robots; Terra produces the data robots learn from and the evaluations that measure their performance.
Who records the robotics data?
Data collection is performed through the Alignerr network and specialist third-party partners using Labelbox-supplied rigs in homes, workplaces, and labs.
Can we commission specific robotics tasks?
Yes, most programs start with a task list and target distribution, and the diversity engine steers collection until coverage is met.
Added by AntSpace from public sources (labelbox.com).