Solutions · VLA training data
VLA training data
VLA training data is paired video and action data — synchronized camera streams, joint states and a task prompt for every episode — from which vision-language-action models such as π0.5 learn to turn instructions into robot actions.
PrismaX supplies it as teleoperated demonstrations on real robot arms, QA-scored episode by episode, from the catalog or collected to your spec.
The data problem
What gets in the way
“Robotics data” isn't defined
Repetitive tasks
Uneven quality
Rigs you can't reproduce
What the data needs
Video and actions, to a standard.
- Video–action pairs of an embodiment completing a semantically meaningful task, with a prompt that matches what happens on screen.
- An overhead camera with full view of the workspace at all times, plus at least one camera on each wrist.
- Every joint angle recorded, so the robot's configuration can be reconstructed, with a URDF whose kinematic chain matches.
- Serial arms with six or seven degrees of freedom and 1-DoF grippers or dexterous hands, suited to cross-embodiment transfer.
- Clean, evenly lit images and smooth, mistake-free demonstrations.
How PrismaX provides it
From the catalog, or collected to spec.
Teleoperation data
Custom data collection
Egocentric data
Industries and scenes
Where the demonstrations come from
- HomeTidying, sorting, folding, placing
- KitchenFood handling and clean-up
- LogisticsPacking, unpacking, bin picking
- IndustrialComponent handling
- OfficeDesk and supply handling
- Lab and ResearchBench work and lab procedures
Related tasks
- Biology LabLab and Research · 108.8k
- Pick and place various fruits into containersKitchen · 2.8k
- Pick and place flowers into vase and vase on shelfHome · 2.0k
- Block manipulationHome · 2.0k
- Pack box of componentsLogistics · 1.8k
- Pick and place slippers into binHome · 1.6k
- Tying shoelacesHome · 1.5k
- Clean up kitchen areaKitchen · 1.4k
Sample videos
See the episodes
A π0.5 policy fine-tuned on 668 YAM pick-and-place episodes picks fruit from a plate, as shown at IROS 2026.
Two robot arms arrange flowers in a vase, seen from the overhead and wrist cameras at the same time.
Related reading
All news
Deep Dive
openpi / π0.5 Fine-Tuning: How-To and Tips
Ablations from a π0.5 fine-tune on YAM pick-and-place data: training scheme, lighting, inference bugs, camera placement and field of view, and backgrounds.
Deep Dive
Latency Versus π0.5 Rollout Performance
A Python pipe buffer added hundreds of milliseconds to our π0.5 rollouts. How we found it, cut inference from 230 ms to 87 ms, and what changed on the robot.
Deep Dive
Quick Note: π0.5 Inference in 16 GB
Two fixes cut our π0.5 rollout stack from about 22 GB to 11.2 GB of VRAM, so it now fits 16 GB cards like the RTX 5080 and 5070 Ti.
FAQ
Frequently asked questions
What is VLA training data?
Data that pairs what a robot sees with what it does: synchronized camera video, joint states and a language prompt describing the task. Vision-language-action models learn from it to map instructions and observations to actions.
Which models has PrismaX trained on this data?
We fine-tune and evaluate π0.5 (openpi) on our own YAM data, and publish what we learn — training schemes, lighting, camera placement, latency and memory — in our deep dives.
How many episodes do I need to fine-tune a VLA?
It depends on the task. As one reference point, our π0.5 pick-and-place fine-tune used 668 episodes, about 10 hours of YAM data, and ran zero-shot on a different rig at IROS 2026.
Does each episode come with a language prompt?
Yes. Every episode is recorded against a task prompt, and an episode whose task doesn't match its prompt fails review.
Can you collect VLA data on our robot?
Tell us the embodiment you train on. The catalog covers YAM, Piper, RLM 63 and TOK2 arms and Enterprise supports 12+ robot arm types; for anything else, describe the embodiment and we scope a custom collection around it.
Next step
Fine-tune on data we train on ourselves.
Tell us the model, the tasks and the robot. We'll point you to the catalog or scope a collection.