VLA-JEPA
Vision-language-action model that pairs a Qwen3-VL-2B backbone with a flow-matching DiT action head, and adds a V-JEPA2 latent video world model as an auxiliary self-supervised loss during training only. At inference only Qwen and the action head run, predicting 7-step action chunks.
open weightsverified 2026-10-24paper →
Architecture
Compute requirements
No sourced requirements yet.
Training requirements: see also the LeRobot compute hardware guide, the authoritative source for LeRobot training envelopes.
Measured inference 0 runs · inference cost only, not task success
No measured runs yet. Submit one with its runtime, precision and camera count.
Robots running VLA-JEPA 8 of 24 with evidence
| Robot | State | Tier | Verified | Detail |
|---|---|---|---|---|
| Franka Research 3 | works with caveats | T1 · upstream documented | 2026-10-24 1d ago | Evidence is the authors' own lab evaluation (100 demonstrations, three pick-and-place tasks); the upstream repo ships training and simulation-evaluation scripts but no FR3 deployment code or FR3 checkpoint. |
| Franka Emika Panda | partial | T1 · upstream documented | 2026-10-25 today | Simulation only: LIBERO (simulated Franka Emika Panda); no real-robot training or inference on a Panda is documented. The paper's real-robot experiments use a Franka Research 3 (see the franka-fr3 cell). |
| I2RT YAM | partial | T3 · community reported | 2026-10-25 today | Training evidenced by public checkpoint(s); on-robot inference not documented by a primary source. |
| ROBOTIS OMX | partial | T3 · community reported | 2026-10-24 1d ago | Training evidenced by public checkpoint(s); on-robot inference not documented by a primary source. |
| OpenArm | partial | T3 · community reported | 2026-10-24 1d ago | Training evidenced by public checkpoint(s); on-robot inference not documented by a primary source. |
| reBot Arm B601-DM | partial | T3 · community reported | 2026-10-24 1d ago | Training evidenced by public checkpoint(s); on-robot inference not documented by a primary source. |
| SO-101 | partial | T3 · community reported | 2026-10-24 1d ago | Training evidenced by public checkpoint(s); on-robot inference not documented by a primary source. |
| Trossen WidowX AI | partial | T3 · community reported | 2026-10-25 today | Training evidenced by public checkpoint(s); on-robot inference not documented by a primary source. |
No verified evidence, but both are available in LeRobot (ecosystem context): AgileX PiPER, ARX5, Bimanual SO-100/SO-101, EarthRover Mini Plus, HopeJR, Koch v1.1, LeKiwi, Reachy 2, SO-100, Unitree G1, Universal Robots UR5e, UFACTORY xArm, XLeRobot.
Framework support
Checkpoints
- lerobot/VLA-JEPA-Pretrain · Converted from ginwind/VLA-JEPA; DROID 1.0.1, 2 exterior views, 7-dim actions.
- lerobot/VLA-JEPA-LIBERO · LIBERO-10; 2 cameras (agentview + wrist), 7-dim actions.
- lerobot/VLA-JEPA-SimplerEnv · OXE Bridge / RT-1; 1 view duplicated x2, 7-dim actions.
- ginwind/VLA-JEPA · Original upstream release (Apache-2.0).
Related terms
Sources
- T1 · upstream documentedofficial-docssource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
LeRobot VLA-JEPA docs: Qwen3-VL backbone, V-JEPA2 video world model, flow-matching DiT action head; world model unused at inference.
“a Vision-Language-Action model that combines a Qwen3-VL language backbone with a self-supervised video world model (V-JEPA2) and a flow-matching DiT action head.”
https://github.com/huggingface/lerobot/blob/v0.6.1/docs/source/vla_jepa.mdx - T1 · upstream documentedofficial-docssource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
Only the Qwen backbone and action head run at inference.
“Only Qwen + the action head are used. The world model is not needed at inference time.”
https://github.com/huggingface/lerobot/blob/v0.6.1/docs/source/vla_jepa.mdx - T1 · upstream documentedofficial-docssource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
Weights follow ginwind/VLA-JEPA (Apache 2.0); integration code Apache 2.0.
“The LeRobot integration code follows the **Apache 2.0 License**.”
https://github.com/huggingface/lerobot/blob/v0.6.1/docs/source/vla_jepa.mdx - T1 · upstream documentedsource-codesource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
Policy type string is `vla_jepa`; default chunk_size 7.
“@PreTrainedConfig.register_subclass("vla_jepa")”
https://github.com/huggingface/lerobot/blob/v0.6.1/src/lerobot/policies/vla_jepa/configuration_vla_jepa.py - T1 · upstream documentedrelease-notessource 2026-07-06verified 2026-10-24 (1d ago)lerobot 0.6.0
VLA-JEPA was added in LeRobot v0.6.0.
“* feat(policies): add VLA-JEPA by @Maximellerbach in https://github.com/huggingface/lerobot/pull/3568”
https://github.com/huggingface/lerobot/releases/tag/v0.6.0 - T3 · community reportedhub-checkpointverified 2026-10-24 (1d ago)
Hub API: lerobot/VLA-JEPA-{Pretrain,LIBERO,SimplerEnv} and ginwind/VLA-JEPA exist, all tagged license:apache-2.0.
https://huggingface.co/lerobot/VLA-JEPA-Pretrain