ABot-ER-9B

Overview

ABot Embodied Reasoning (ABot-ER) is a unified embodied reasoning model family developed by AMAP CV Lab, Alibaba Group. It brings perception, planning, and execution-related reasoning into a shared vision-language model. This repository provides the ABot-ER-9B weights and accompanying model-loading code.

The model uses the Task-Conditioned Adaptive Training Strategy described in the technical report. Task tokens select lightweight, sequence-level low-rank residuals in selected query and value projections, while the language backbone remains shared across tasks.

Capabilities

• Spatial & Relational Reasoning: object positions, distances, directions, and relationships.
• Scene & State Understanding: objects, attributes, states, and functions in a scene.
• Temporal & Event Reasoning: actions, event order, changes, and outcomes in video.
• Point Grounding & Affordance: instruction-conditioned localization of objects and regions for interaction or placement.
• Long-Horizon Action Organization: action-plan generation conditioned on task goals and environmental information.

The model generates answers, coordinates, and plans. Executing actions requires an agent runtime or downstream controller.

Model information

Item	Description
Model	ABot-ER-9B
Team	AMAP CV Lab, Alibaba Group
Backbone	Qwen3.5-9B
Training	One Stage A cross-dimension foundation run, followed by one Stage B joint post-training run from the selected Stage A checkpoint
Method	Task-Conditioned Adaptive Training Strategy
Intended use	Embodied AI research, multimodal understanding, and task-planning evaluation

Both stages use supervised fine-tuning; Stage B is not reinforcement learning. See USAGE.md for the released checkpoint's residual layers, task routing, and loading requirements.

Download and use

• Project code and documentation
• Model loading and task routing
• Training data
• ModelScope mirror

Install huggingface_hub and set HF_REVISION to an immutable commit SHA from this repository.

import os
from huggingface_hub import snapshot_download

model_dir = snapshot_download(
    repo_id="acvlab/ABot-ER-9B",
    revision=os.environ["HF_REVISION"],
)

Use the included loading implementation and select the task route as documented in USAGE.md. Loading the weights as an ordinary backbone without the task-conditioned implementation is unsupported.

This model is intended for research and evaluation. Generated outputs require application-specific validation and must not be treated as a guarantee of safe real-world robot execution.

License and provenance

See LICENSE for backbone and component terms. This repository does not relicense all underlying model weights under a single Apache or MIT license. Applicable upstream terms remain in force.

Citation

ABot-ER: Unified Embodied Reasoning through Agentic Data-Model Co-Evolution.

@misc{ding2026aboterunifiedembodiedreasoning,
  title={ABot-ER: Unified Embodied Reasoning through Agentic Data-Model Co-Evolution},
  author={Qiuyu Ding and Heng-Da Xu and Changda Xia and Jiajun Shen and Zihao Guan and Zhengchen Yuan and Jianfang Zhang and Di Yang and Jiayi Tian and Yuting Xu and Fei Wang and Shiao Liu and Wei Zhang and Xiaolong Wu and Wenbin Tang and Zedong Chu and Mu Xu},
  year={2026},
  eprint={2610.18355},
  archivePrefix={arXiv},
  primaryClass={cs.AI},
  url={https://arxiv.org/abs/2610.18355}
}