Juno: Taming Predictive Latents for Vision-Language-Action Models

Yuchen Zhu, Chenyi Xu, Yulin Zhang, Gang Xu, Wentao Zhu

---

Juno studies how predictive visual latents can help a vision-language-action (VLA) policy reason about the consequences of its actions. The method aligns future visual representations with the control stream, while a mixture-of-transformers action branch keeps language-conditioned reasoning and action prediction coupled.

The release is organized as a small, reproducible repository. It contains the Juno training stack, the standalone juno-policy model, a Bridge/Fractal configuration, and a launcher whose paths are supplied through environment variables.

News

[2026/10] 🚀 The paper is available on arXiv.

[2026/10] 🚀 Initial standalone release of the Juno Bridge/Fractal policy-learning recipe.

TODO

• Juno Model Architecture
• Training Code
• Checkpoints Release
• JEPA Training Code
• TTT Code

Overview and Key Features

Overview of Juno. Future-frame JEPA targets are aligned with the control stream of a vision-language-action policy, while a mixture-of-transformers reasoning branch keeps language-conditioned prediction coupled to action generation.

Training pipeline

Training pipeline of the released Bridge/Fractal policy-learning stage. Dataset and checkpoint paths are runtime inputs; the launcher records the resolved config and run directory through the Juno trainer.

What is released

This repository releases the Bridge/Fractal policy-learning stage used to train the juno-policy model.

---

🎒 Quick Start

📖 First run? Clone the repo, install the training stack, then point BASE_VLM, OXE_DATA_ROOT, and JEPA_CKPT_PATH at assets kept outside this repository. See docs/reproducibility.md for the exact recipe.

1. Clone and install

The training environment follows the same conda recipe as StarVLA (python=3.10), with the environment renamed to Juno.

git clone https://github.com/juno-policy/Juno.git
cd Juno

conda create -n Juno python=3.10 -y
conda activate Juno

pip install -r requirements.txt
pip install flash-attn --no-build-isolation
pip install -e .

This stack matches a working StarVLA environment: Python 3.10, PyTorch 2.8 + CUDA 12.8, transformers==4.57.0, accelerate==1.5.2, deepspeed==0.16.9, and peft==0.17.1. Install a CUDA-compatible PyTorch build for your cluster if pip install -r requirements.txt does not already pull one.

⚠️ Common Issues

flash-attn must match your CUDA toolkit (nvcc) and PyTorch versions. The --no-build-isolation flag resolves most issues. Check your environment:

nvcc -V
pip list | grep -E 'torch|transformers|flash-attn'

We have verified that flash-attn==2.7.4.post1 works well with nvcc 12.0 / 12.4. On CUDA 12.8 clusters, pick a flash-attn wheel that matches PyTorch 2.8.

2. Prepare external assets

Keep datasets and checkpoints outside this repository. Set:

export BASE_VLM=/path/to/Qwen3-VL-4B-Instruct
export OXE_DATA_ROOT=/path/to/bridge_fractal_lerobot
export JEPA_CKPT_PATH=/path/to/jepa_encoder.pt
# Optional: external le-wm package and experiment settings
export LEWM_ROOT=/path/to/le-wm
export NUM_PROCESSES=8
export WANDB_PROJECT=juno-policy

BASE_VLM, OXE_DATA_ROOT, and JEPA_CKPT_PATH are required. After conda activate Juno, the launcher uses that environment's python, writes runs under runs/, and supports PYTHON_BIN, CONFIG_PATH, RUN_ROOT_DIR, RUN_ID, IS_RESUME, WANDB_MODE, and WANDB_ENTITY overrides.

3. Validate and train

conda activate Juno
bash scripts/validate.sh
bash scripts/run_bridge_fractal_jepa7_mot_train.sh

The command uses DeepSpeed ZeRO-2 and the launch settings in juno/config/deepseeds/deepspeed_zero2.yaml. Start with a small NUM_PROCESSES and an offline W&B run when checking a new cluster.

---

Benchmark Results

The numbers below are reported paper results, not a benchmark run generated by this repository. SimplerEnv entries are means over 4 evaluation runs. RoboCasa-GR1 uses 50 episodes per task. Each real-robot condition uses 20 trials. The best result in each column is in bold. See the paper for task definitions and evaluation details.

SimplerEnv (WidowX)

Success rate (%). Arrows are the change from Qwen3GR00T.

Method	Average	Spoon on Towel	Carrot on Plate	Stack Green Block	Eggplant in Basket
OpenVLA-OFT	41.8	34.2	30.0	30.0	72.5
RoboVLM	42.7	50.0	37.5	0.0	83.3
Magma	44.8	37.5	29.2	20.8	91.7
CogACT	51.3	71.7	50.8	15.0	67.5
SpatialVLA	34.4	20.8	20.8	25.0	70.8
TraceVLA	27.7	12.5	16.6	16.6	65.0
VideoVLA	53.1	75.0	20.8	45.8	70.8
VLA-JEPA	57.3	75.0	70.8	12.5	70.8
π0	53.1	29.2	62.5	29.2	91.6
π0.5	57.1	49.3	64.7	44.7	69.7
Isaac-GR00T-N1.6-Bridge	57.1	64.5	65.5	5.5	93.0
Qwen3GR00T	60.9	72.9	60.4	14.6	95.8
Juno	68.5 (↑7.6)	86.5 (↑13.6)	59.4 (↓1.0)	31.3 (↑16.7)	96.9 (↑1.1)
Juno-TTT	72.7 (↑11.8)	94.8 (↑21.9)	62.5 (↑2.1)	37.5 (↑22.9)	95.8

RoboCasa-GR1 Tabletop

Unweighted mean success rate (%) within each task group. Average is over all 24 tasks.

Method	PnP + Close	Cuttingboard	Placemat	Plate	Tray	Avg.
GR00T-N1.6	24.2	56.9	51.9	57.6	55.1	47.6
Qwen3PI	42.3	46.0	43.5	44.0	44.0	43.9
Qwen3OFT	43.7	50.4	41.5	61.0	49.2	48.8
Qwen3FAST	35.0	50.4	33.5	45.0	32.0	39.0
Qwen3GR00T	50.3	52.8	38.0	58.5	39.2	47.8
Juno	57.3	64.8	55.5	68.0	53.6	59.6

Real robot, frozen policy (ALOHA right arm)

Success rate (%) over 20 trials. Shifts are cumulative: background, then height, then object.

Method	In-Domain	Background	+ Height	+ Height + Object
Qwen3GR00T	40.0 (8/20)	0.0 (0/20)	0.0 (0/20)	0.0 (0/20)
Juno	100.0 (20/20)	75.0 (15/20)	70.0 (14/20)	70.0 (14/20)

Real robot, test-time training

Success rate (%) over 20 trials. Juno-TTT is adapted separately in each condition.

Method	Gaussian Noise	Dynamic Lighting
Juno (frozen)	40.0 (8/20)	55.0 (11/20)
Juno-TTT	65.0 (13/20)	70.0 (14/20)

---

Repository layout

Juno/
├── assets/                         # Icon and paper figures
├── configs/                        # Reproducible training configuration
├── deployment/                     # Image preprocessing helpers used by Juno
├── docs/                           # Reproducibility and source map notes
├── scripts/                        # Validation and training entry points
├── juno/
│   ├── model/framework/VLM4A/      # QwenGR00T base + juno_policy.py
│   ├── model/modules/world_model/  # JEPA modules
│   ├── dataloader/                 # LeRobot mixture and transforms
│   └── training/                   # Juno trainer
├── CITATION.cff
├── LICENSE
└── requirements.txt

• Dataset and checkpoint paths are runtime inputs; no private machine paths are required.
• docs/source_map.md maps the paper concepts to the implementation files.
• docs/reproducibility.md lists the exact recipe and the differences between this policy stage and the full paper system.

Cite Juno

If you use Juno, please cite the paper:

@misc{zhu2026junotamingpredictivelatents,
      title={Juno: Taming Predictive Latents for Vision-Language-Action Models}, 
      author={Yuchen Zhu and Chenyi Xu and Yulin Zhang and Gang Xu and Wentao Zhu},
      year={2026},
      eprint={2610.09940},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2610.09940}, 
}

Acknowledgements

This project draws inspiration and references from several notable open-source initiatives, including:

• StarVLA
• Qwen-VL
• LeRobot
• GR00T
• DeepSpeed

Please follow the upstream licenses and model terms when redistributing weights or datasets.