The desktop app

One window.
Every step.

Record, label, train, infer, fork, publish — all behind a single sidebar. No half-dozen browser tabs. The seam between cloud and local is invisible.

The app

One window. Every step.

Record, label, train, infer, fork, publish — all behind a single sidebar. Click through the workspace. Compatibility checks, latency budgets, and per-backbone envs are baked in.

axion · my_pen_pickup_v3 INFER
Overview
SO-101 · wrist cam · calib 8f2c… · seoul region
Phase 2-A Online
Episodes
142
+ 12 this week
Models
7· 3 verified
last: 2h ago
Storage used
12.4GB
of 50 GB plan
Local GPU
24GB · idle
desktop · ready
Recent activity
2h ago
Trained my_pen_pickup_v3cloud GPU · 32 min
success
5h ago
Collected session_1210 episodes · wrist cam
data
yesterday
Forked axion-base-so101-grootn17-v0.4parent: nvidia/GR00T-N1.7-3B
fork
2d ago
Published eraser_v2 to marketplacegrade · 🟢 new · public
market
conda · axion_grootn17 cached base · v0.4 loaded cloud sync · ✓ idle all systems nominal
Collect · session_12
AXION capture · wrist + topview · 30 FPS · 1920×1080
EP 7 / 10 Recording
CAM_0 · wrist
CAM_1 · topview
Prompt · this session
pick up the eraser and place it on the right
Episodes captured
7 done · 1 recording · 2 queued
parquet · 142MB h264 · 1.2GB auto-meta ✓ teleop · stable
Train · my_pen_pickup_v3
AXION cloud training · container axion-pi05:1.0 · pre-emption safe
cloud GPU Epoch 8 / 15
Backbone
4 available · pick one
GR00T N1.7 nvidia · 3B
First-class baseline. exp11 v3 skill · 1.66° open-loop MAE.
live
Pi 0.5 physical intelligence
Better long-horizon manipulation on this dataset. Skill-build + cloud inference verified.
★ recommended live
OpenVLA · 7B
Largest model. Pairs with cloud inference once Phase 2-C lands.
queued · soon
AXION-S0 axion · SO-101 dedicated
Native backbone trained on the SO-101 corpus. Faster, smaller, calibration-aware.
queued
Base
axion-base-so101-pi05-wrist-v0.2 2400ep refluxed · 2026-05-21
Method
Standard skill build strength 16 · lr 5e-5 · 15 epochs
Loss
0.0421↓ -14% vs prev epoch
epoch 0now
Joint MAE
2.14°↓ approaching 1.66° target
epoch 0now
cloud GPU · pre-emption safe checkpoints · cloud storage ✓ ETA · 1h 22m on track
Infer · my_pen_pickup_v3
on axion-base-so101-grootn17-wrist-v0.4 · skill build
Local GPU Running
CAM_0 · wrist
OVERLAY · action
Latency · p50
18ms
Step
142 / chunk 16
compat ✓ camera compat ✓ robot compat ✓ calib 8f2c… safety nominal
Fork tree · so101 · wrist
7 skills · 2 verified · 1 yours · branch from any node
Layout · graph HEAD · my_pen_pickup_v3
upstream axion base community skills your forks
nvidia/GR00T-N1.7-3B
#a3f7c1
upstream 3Bparams
axion-base-so101-v0.4
#8f2c4a
base 3000ep refluxed
eraser_pickup_v3
#d12e9b
verified 18 340dl
pen_pickup_v2
#5a91f0
verified 7 120dl
colour_sort_v1
#7c3aed
new 2 34dl
eraser_top_corner
#2bd1a8
new 0
eraser_with_distractor
#f0a23c
new 0
my_pen_pickup_v3
#e8b34d
you 0 2h ago
parent · pen_pickup_v2 siblings · 2 ahead · 1 commit click any node to fork / branch
Marketplace
filtered to · so-101 · wrist · grootn17
142 skills Sort · recent
eraser_pickup_v3
verified
axion-team · 2d ago
18forks340dl1.66°MAE
pen_pickup_v2
verified
axion-team · 5d ago
7forks120dl2.10°MAE
colour_sort_v1
tested
user_c · 1d ago
2forks34dl3.40°MAE
eraser_top_corner
new
user_a · 3h ago
0forks4dl
pour_liquid_steady
tested
lab_kr · 4d ago
5forks89dl2.55°MAE
stack_blocks_3
new
user_d · 6h ago
1forks11dl
filtered by compat · 142 verified · 12 tested · 38 click card · open detail
Multi-backbone

One pipeline. Any foundation model.

AXION is built backbone-agnostic from day one. Same dataset, same UI, same fork tree — pick the foundation model that fits your task. New backbones add via yaml, not code.

01
GR00T N1.7 nvidia · 3B
First-class baseline. exp11 v3 skill — 1.66° open-loop MAE, 40% real-robot success.
Live · Phase 2-A
02
Pi 0.5 physical intelligence
Better long-horizon manipulation on this dataset. Skill-build + cloud inference verified.
Live · Phase 2-A
★ Recommended
03
OpenVLA · 7B
Bigger model, bigger GPU. Paired with cloud inference at Phase 2-C.
Queued · soon
04
AXION-S0 axion · SO-101 dedicated
Native backbone trained on the SO-101 open data corpus. Faster, smaller, calibration-aware.
Queued
05
Your backbone · bring a yaml
Implement the AXION backbone interface. Drop a yaml in config/backbones/. Done.
User-defined
config/backbones/grootn17.yaml

New backbone = new yaml. Zero code changes.

Every backbone declares its env, weights, runner, and what it supports. Add a yaml, AXION enumerates it in the picker, validates compatibility, and pulls a conda-pack tarball on demand.

name: grootn17
display: "GR00T N1.7"
version: "N1.7-3B"
conda_env: axion_grootn17
runner: axion.training.runners.grootn17

supports:
  skill_types: ["standard", "advanced"]
  cameras: ["wrist", "topview"]
  os: ["linux", "windows-wsl2"]

default_config:
  skill: { strength: 16, lr: 5e-5, epochs: 15 }
# jayden builds the env once. conda-pack uploads
# a tarball. user downloads it. no pip wars.
Cloud + local hybrid

Cloud trains it. Your robot runs it.

Training is heavy and intermittent — perfect for managed cloud GPU. Inference is real-time and continuous — perfect for the GPU already on your desk. AXION orchestrates both.

01 · YOUR DESK

Captured episodes

Teleoperated demos, captured locally in the AXION app.

parquet + h264 auto-meta + calib hash mandatory open*
upload
02 · AXION CLOUD

Managed cloud skill build

Container image axion-pi05:1.0 on a managed GPU node. 2–7 hours per skill. Pre-emption safe.

cloud GPU · 24GB class pre-emption safe cloud-stored checkpoints
skill
03 · YOUR GPU

Local 50ms control

Your skill hot-swaps onto the cached base. Control loop stays under 50ms, frames never leave your network.

skill hot-swap safety stop · local cloud infer · phase 2-C
Avg training time
2–7hper skill build
Control-loop latency
< 50mslocal · Phase 2-A
Cold-start (cloud infer)
~80msSeoul region · cloud infer
Supported hardware Request early access