Physical AI
Fine-tune and deploy Vision–Language–Action models for real robots
powered by Positronic Robotics
Quick start
Fine-tune a VLA model
Pick a base model, point at your dataset, train on Nebius H100s. Deploy in one click.
Deploy an inference endpoint
Serve any open-source VLA on serverless GPUs. Robots connect over HTTPS, scale to zero when idle.
Manage datasets
Upload teleop data, browse the public catalog, inspect any episode in 3D.
Active endpoints
View all →
| Name | Model | Status | P50 latency | Req / min | Replicas | Today | |
|---|---|---|---|---|---|---|---|
|
warehouse-pick-prod
ep-9k2a · OpenPI π0.5
|
OpenPI π0.5 | Active |
38 ms | 1,420 | 2 / 4 | $12.84 |
|
|
fold-shirt-bimanual
ep-7r3b · GR00T N1.6
|
GR00T N1.6 | Active |
44 ms | 312 | 1 / 2 | $3.62 |
|
|
cable-insertion-staging
ep-2m8c · SmolVLA
|
SmolVLA | Idle |
— | 0 | 0 / 1 | $0.18 |
|
Recent fine-tuning runs
View all →
| Name | Base model | Dataset | Status | Hardware | Duration | Cost |
|---|---|---|---|---|---|---|
warehouse-pick-v3 run-a8df21e2c |
OpenPI π0.5 | warehouse-bins-12k | Active |
1× H100 NVLink | 14m 32s / 50m | $0.98 |
fold-shirt-bimanual run-72bcef901 |
GR00T N1.6 | laundry-bimanual-v2 | Complete |
1× H100 NVLink | 52m 18s | $3.42 |
kitting-tray run-5e2ab98c1 |
ACT | kitting-pilot-300 | Failed |
1× L40S | 4m 11s | $0.09 |
My datasets
View all →
| Name | Duration | Episodes | Size | Used by |
|---|---|---|---|---|
warehouse-bins-12k pick-and-place · 30 Hz |
110h 18m | 12,418 | 18.4 GB | 3 runs |
| 112h 48m | 4,512 | 31.2 GB | 1 run | |
usb-insertion-pilot cable-into-port · 60 Hz |
6h 48m | 980 | 2.1 GB | 1 run |
Active endpoints
3
2 live · 1 idle
Fine-tuning runs (mo)
42
128 GPU-hours · ▲ 18%
Datasets
5
303h 31m · 64.7 GB
Spend (mo)
$481
of $2,000 budget
