Native LeRobot and RLDS output. Multi-layer annotation pipeline. GDPR-compliant by design. Built in Europe for robotics teams scaling from demos to deployment.
The problem
Imitation learning, behavioral cloning and VLA training are bottlenecked by the same three gaps: the egocentric viewpoint, the annotation stack a policy actually consumes, and the compliance posture you need to deploy in Europe. EgoVista is built to close all three.
Real-world data
Manipulation policies only generalize as far as their training data reaches. A model that has only seen one kitchen struggles in the next one. These frames come from real homes and workshops across Europe, with different lighting, layouts, and contributors. The variety is the point: it is what lets a policy hold up outside the room it was trained in.
How it ships
Whether you start from scratch or already have footage, we deliver training-ready datasets in LeRobot or RLDS format.
Tell us what task your robot needs to learn, the environment, and the objects involved. We're building a contributor network to film targeted egocentric footage on demand. Custom collection projects open Q3 2026. Contact us to discuss your needs and get early access.
Best for: teams starting from scratch, specific environments needed, diverse contributors
Pricing: TBD
Discuss your project →Already have egocentric or manipulation video? Send it to us. Our pipeline adds 7 annotation layers and exports in LeRobot or RLDS format. Sample batches in days, production runs in weeks.
Best for: teams with existing footage, teleop recordings, in-house captures
Pricing tailored to your project scope
Send your videos →Annotation service · how it works
Send us a link to your files or email them to contact@egovista.app. We accept any format, any size.
Pick the annotation modules you need: pose, segmentation, depth, actions, contacts.
Receive annotated data in LeRobot, RLDS, or custom format. Sample batches in days, production runs in weeks.
Hand pose annotation, depth estimation, EgoHOS segmentation, contact timing, action labels, packaged together in a single LeRobot or RLDS dataset so your training loop only learns one IO path. Mix and match modules per project.
MediaPipe: 21 keypoints per hand per frame
MediaPipe + Depth: 21 keypoints × 2 hands in 3D space
Depth Anything: per-frame depth maps
EgoHOS: 5 classes per frame, EU GPU inference
Hand-object contact transitions, per frame
Gemini Vision (EU): timestamped natural-language labels
MediaPipe Face Detection: applied before any external call
Focal length, principal point, distortion, per clip
Episode boundaries, schema versioning, QA score per clip
Need a module that's not listed? Ask. We ship new modules in 2 weeks on request.
Datasets ship in the LeRobot v3.0 format out of the box. Load them straight into the Hugging Face LeRobot stack without writing a single conversion script. Tested against the canonical LeRobot dataset spec.
Need RLDS? We export to the TFDS-backed schema used by Open X-Embodiment so your dataset slots into any existing imitation learning or VLA training pipeline, no glue code required.
EU-first by design
Face anonymization before any external processing, EU-only infrastructure, documented legal bases per processing activity, EU AI Act-ready audit trail. GDPR compliance details.
Faces blurred at source before any annotation. DPA available, contributor consent collected per mission, RoPA documented.
Storage Cloudflare R2 EU, GPU RunPod Amsterdam, Vertex AI europe-west4. No data crosses the Atlantic.
Network of European contributors recording in EU jurisdictions. Right to erasure, opt-out, and full data lineage on demand.
Every clip, every annotation pass, every model invocation logged with timestamps. Inspect-ready for EU AI Act compliance.
Quality assurance
Annotation quality is checked per clip, scored, and shipped with the dataset. You see exactly what was reviewed, what passed, and where uncertainty remains.
Schema validation, hand pose accuracy, action label precision and segmentation coverage measured per clip. Average score across delivered datasets sits above 90%.
Labels are model predictions, not ground truth. Every dataset includes per-frame uncertainty metadata so your team knows what to trust and what to filter before training.
Every annotation pass logs the model version, timestamp, and EU region. If your downstream training surfaces a regression, you can trace it back to the exact pipeline run.
Use cases
The same annotated egocentric video powers three distinct training workflows. Pick the one closest to your stack, or talk to us about a hybrid.
Train BC, offline RL and VLA policies on egocentric human demonstrations annotated with hand pose, contact timing, depth, and action labels, shipped in your training format.
manipulation policy training →Dense, diverse human demonstration data with the annotation layers that imitation learning baselines actually consume, no custom preprocessing required.
imitation learning workflows →Anonymized, GDPR-clean egocentric video at scale for pretraining vision encoders, action models, and large-scale world models, with full provenance.
Talk to us about pretraining →How it works
No black box. You see exactly what you brief, what we ship, and how it was annotated.
You specify the task, target environment, format (LeRobot or RLDS), schema, and any compliance constraints.
Our European contributor network captures egocentric footage; the 9-layer pipeline annotates, anonymizes, and packages it under EU-only infra.
Encrypted signed URL, your chosen format and schema, with a per-dataset QA report and full annotation provenance.
FAQ
Most modern container/codec combinations: MP4 (H.264, H.265), MOV, MKV, WebM. Resolution from 720p to 4K. We strongly prefer egocentric footage (head-mounted camera, smart glasses, or chest-mounted GoPro), but we can work with first-person teleop or wrist-camera footage if that is what your project requires.
Timeline depends on volume and engagement model. For annotation-only on a small batch of existing video, expect days. For full collection + annotation, expect two to six weeks for a first pilot batch, depending on task complexity and contributor matching. We always agree on a delivery date before kickoff and ship the QA report with the data.
Three layers: (1) face anonymization applied locally with MediaPipe before any external API call, so no identifiable face data ever crosses an external boundary; (2) EU-only infrastructure for storage, GPU inference, and labeling (Cloudflare R2 EU, RunPod Amsterdam, Vertex AI europe-west4); (3) contributor consent collected per mission with documented legal bases and a right to withdrawal that we honor within thirty days.
Yes. Tell us the task, the environment, and any constraints on objects, lighting, or contributor profile. We brief our European contributor network, match the closest profiles, and ship targeted clips for that exact scenario. If you need a task that has never been captured, we can also stage controlled recordings.
LeRobot v3.0 is the Hugging Face robotics dataset format: parquet shards plus MP4 frames plus JSONL metadata, native to the LeRobot training stack. RLDS is the TFRecord-based format used by Open X-Embodiment and the broader TFDS ecosystem. We ship both from the same source data; pick whichever matches your training pipeline.
Annotations. Our hand pose, depth, segmentation, and action labels are model predictions, not ground truth. They have measurable accuracy but are not human-verified frame by frame. Every dataset ships with uncertainty metadata so your team can filter or weight noisy samples. Human-verified ground truth is available on request for evaluation subsets.
The QA report ships before the dataset is final. If the score is below the threshold we agreed at kickoff, we re-annotate or recollect the affected clips at no additional cost. We also support evaluation pilots (a small sample before scaling) so you can validate quality on your own benchmark.
Yes. We have a standard mutual NDA template that covers your project specifics, training objectives, and the contents of the dataset. We can sign it before the first technical call. Email contact@egovista.app and we can have it back to you within 24 hours.
Tell us the task, the environment, and the format. We'll ship you a fully annotated egocentric video sample in LeRobot or RLDS so your team can evaluate quality before scaling.
No commitment · 48h response · NDA available