InternVLA-A1.5-RAB

InternVLA-A1.5 fine-tuned for RobotAgentBench. This release contains checkpoint 120000, its matching normalization statistics and control schema, and the Qwen3.5-2B VLM files required by the inference pipeline.

Download

pip install -U huggingface_hub
hf download AetherLabs-AI/InternVLA-A1.5-RAB --local-dir weights/InternVLA-A1.5-RAB
cd weights/InternVLA-A1.5-RAB
shasum -a 256 -c SHA256SUMS

The complete download is approximately 10 GB. The policy checkpoint is 5.39 GB; the vlm/ directory includes the matching VLM weights, processor and tokenizer. No training data, optimizer state or simulation assets are included.

Run with RobotAgentBench

Use the InternVLA example and installation guide. After downloading, use these absolute local paths in internvla configure:

Argument	Path relative to the downloaded model directory
--checkpoint	.
--stats	stats.json
--stats-key	G1_omnipicker (key, not a path)
--control-schema	control_schema.json
--vlm	vlm/

For example, append the following model arguments to the full configure command in the linked guide, using your own absolute download directory:

--checkpoint /absolute/path/to/weights/InternVLA-A1.5-RAB \
--stats /absolute/path/to/weights/InternVLA-A1.5-RAB/stats.json \
--stats-key G1_omnipicker \
--control-schema /absolute/path/to/weights/InternVLA-A1.5-RAB/control_schema.json \
--vlm /absolute/path/to/weights/InternVLA-A1.5-RAB/vlm

Isaac Sim, the benchmark assets, a prepared evaluation instance and the separate model Python environment are installed/configured as described in the guide. Downloading the model does not install or start the simulator.

Model interface

• Inputs: task instruction, three RGB camera views and 23-dimensional measured robot state.
• Output: 50-step action chunks, with 22 supervised physical channels padded to 32 in model space.
• Actions: 20 absolute joint targets, forward base velocity and yaw rate; evaluation control is 20 Hz.
• Use the included statistics for the original mean/std decoding and the benchmark's action adapter.
• The runner selects optimized, action-only BF16 inference and the explicitly supplied local VLM path.

The published weight bytes are unchanged. Deployment-only absolute paths were removed from config.json; private provenance fields were removed from the control schema without changing its channel order, units or control semantics. The VLM files are retained unchanged. This is a robot-specific checkpoint, not a general-purpose chat model or a drop-in policy for other robot embodiments. Do not deploy on physical hardware without appropriate safety validation.

Attribution and license

Fine-tuned for RobotAgentBench by AetherLabs-AI. Based on InternVLA-A-series from Shanghai AI Laboratory, source revision e6fc904f9edbfb14532e97095fc2372202517f76. The InternVLA-derived release is distributed under CC BY-NC-SA 4.0. The separately bundled Qwen3.5-2B files in vlm/ retain their Apache 2.0 license (vlm/LICENSE). See LICENSE and NOTICE.md.

Citation

@misc{mao2026robotagentbenchopenworldrobotautonomy,
      title={RobotAgentBench: Towards Open-world Robot Autonomy}, 
      author={Lingjun Mao and Jinglin Cao and Lukun He and Wenpeng Xu and Yuchen Yan and Hengyu Jin and Zijun Zhang and Jinzhou Tang and Songyao Jin and Ziche Liu and Yifei Shao and Junbo Huang and Fang Nan and Ruobing Han and Ziqiao Xi and Ziming Xu and Shuang Liang and Sibo Zhu and Wenyi Wu and Xinyue Wang and Yiming Huang and Xinkai Zou and Gao Zhu and Kun Zhou and Biwei Huang},
      year={2026},
      eprint={2610.34649},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2610.34649}, 
}