Tell us what your robot needs to learn.
Share your task and timeline. We will come back with a collection plan and a quote.
On-demand training data for physical AI.
Abundance collects, annotates, and delivers training data your models need to learn from the physical world, so you can focus on building the future.
01 · Request
Task
Household manipulation
Modalities
Environments
Real homes, varied lighting
Delivery
Rolling weekly batches
02 · Capture
Folding laundry
Bedroom · Session A
Loading dishwasher
Kitchen · Session B
Sorting groceries
Pantry · Session C
Wiping surfaces
Kitchen · Session D
03 · Deliver
household_manip_v1/
episodes/
ep_0001.mcap
ep_0002.mcap
annotations/
actions.parquet
metadata.json
Language models learned from the internet. Robots need data that doesn’t exist yet — so we go out and collect it.
What we capture
Every modality your model learns from.
From first-person video to full teleoperation trajectories, captured by trained contributors in real environments.
Egocentric video
First-person footage of people doing real tasks, with synchronized depth and hand-pose tracking.
Teleoperation demonstrations
Expert-driven robot trajectories with joint states, actions, and camera streams for imitation learning.
Motion capture
Full-body and dexterous hand motion for humanoid control, retargeting, and whole-body policies.
Multimodal sensor data
RGB-D (color and depth), LiDAR (Light Detection and Ranging), IMU (Inertial Measurement Unit), audio, and tactile streams — time-aligned.
How it works
From spec to training-ready in three steps.
01
Specify
Tell us the task, embodiment, environments, and volume. We scope a collection plan with you.
02
Capture
Our contributor network records in real-world settings, with calibrated hardware and consistent protocols.
03
Deliver
Annotated, quality-checked episodes arrive in the format your training stack already reads.
Built for
Teams teaching machines to act.
Humanoid robotics
Whole-body and bimanual data for general-purpose humanoids working in homes, warehouses, and factories.
Dexterous manipulation
Fine-grained grasping, tool use, and contact-rich tasks with hand pose and tactile signals.
Robot foundation models
Diverse, large-scale episodes across tasks and environments for VLA (Vision-Language-Action) pretraining.
Evaluation & benchmarking
Held-out, real-world test sets to measure how policies generalize before they ship.
Consent by design
Every contributor and location opts in, with clear licensing for commercial model training.
Privacy protected
Faces and PII (Personally Identifiable Information) are blurred or removed before anything leaves our pipeline.
Quality you can audit
Multi-stage QA (Quality Assurance) with per-episode metadata, so you know exactly what you are training on.