# AR-WAM

> Source: https://aiwiki.ai/wiki/ar_wam
> Updated: 2026-10-20
> Categories: AI Research, Robotics, World Models
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "AR-WAM." aiwiki.ai, 20 Oct 2026. https://aiwiki.ai/wiki/ar_wam
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

AR-WAM is a research system for [robotic manipulation](https://aiwiki.ai/wiki/robotics) that connects a high-level [agent](https://aiwiki.ai/wiki/agent) to a compact [world action model](https://aiwiki.ai/wiki/world_action_model). The agent identifies a target with a bounding box and selects a learned operation, such as picking or placing.[3]

Yicheng Jiang, Zesen Gan and colleagues introduced it in a September 20, 2026 preprint, revised October 2. The policy has no language encoder; the external agent handles language and planning.[1]

## Architecture

The 0.5-billion-parameter policy combines a frozen 0.3-billion-parameter visual encoder with a shared 0.2-billion-parameter transformer.[2] The implementation uses [DINOv3](https://aiwiki.ai/wiki/dinov3) for visual features. A reasoning expert predicts the target's mask, terminal end-effector poses and a structured task description. A world action expert generates actions while learning two future targets: a compact representation of visual change and a semantic representation from the reasoning backbone's future-observation output.[3]

The model predicts future representations rather than requiring pixel-space video generation. Its conditioning combines operation embeddings and visual prompts through adaptive layer normalization. The reasoning output makes the policy's intended target and motion available for inspection.[3]

## Agent interface

The interface separates three operations:[1]

| Primitive | Responsibility |
| --- | --- |
| `detect` | The agent grounds its target in the image and returns a box. |
| `execute` | The policy performs the selected operation until a trigger event. |
| `query` | Execution feedback returns to the agent. |

Triggers include gripper opening or closure, accumulated end-effector displacement and a timeout. The agent retains interaction history and retries failed operations; the policy has no recurrent memory.[1]

## Data and implementation

The released RoboTwin dataset contains 27,500 episodes across 50 tasks, divided into 2,500 clean and 25,000 randomized episodes, with 7,269,561 frames. Its schema includes 23 skill IDs and five semantic style IDs. Episode files contain images, robot state, actions, object boxes and event annotations. The dataset card labels the release MIT-licensed.[4]

This 50-task release is broader than the ten-task RoboTwin evaluation reported on the project page.[2][4]

The published training configurations specify the following settings:[5][6]

| Setting | Configuration |
| --- | --- |
| Action/state dimensions | 14/16 |
| Action chunk | 32 steps |
| Camera | `cam_head` |
| Configured image size | `[320, 240]` |
| Global batch size | 128 |
| First policy stage | 550,000 updates; learning rate 0.0002 |
| Second policy stage | 200,000 updates; learning rate 0.00004 |

The repository provides MIT-licensed code. Its instructions pretrain a foresight-gist extractor for 150,000 iterations before policy training. Stage two starts from stage-one weights with a fresh optimizer and scheduler. The documented reference environment uses Python 3.10, PyTorch 2.7.1 and CUDA 12.8; its example commands use eight GPUs.[3]

## Reported evaluations

The authors report the following success rates. RoboTwin uses 50 trials per task and setting. RMBench also uses 50 trials per task; AR-Q and AR-K pair the policy with Qwen3.8-27B and Kimi-K3, respectively.[2]

| Evaluation | AR-WAM | Comparison |
| --- | --- | --- |
| RoboTwin 2.0, clean | 87.2% | Fast-WAM: 87.8% |
| RoboTwin 2.0, randomized | 84.2% | Fast-WAM: 82.5% |
| Both RoboTwin settings, average | 85.7% | Not the clean-only result |
| RMBench, all tasks | AR-Q: 67.8%; AR-K: 51.1% | Cortex: 61.9% |
| RMBench, memory-heavy M(n) tasks | AR-Q: 63.5%; AR-K: 52.5% | Goal2Skill: 38.7% |

Real-robot tests used [Astribot S1](https://aiwiki.ai/wiki/astribot_s1), 100 demonstrations for fine-tuning per task and 20 trials per task. Success was 70% for repeated bell pressing, 65% for picking a specified bottle twice and 80% for placing an egg in a specified cell. The 71.7% average exceeded pi0.5's 35.0% by 36.7 percentage points. This was not zero-shot transfer.[2]

## Latency and limitations

On one 24 GB RTX 4090, with bf16 precision, identical inputs and warm-up runs, the authors measured 14.09 ms per action chunk. This policy measurement should not be read as complete agent-to-robot latency.[1]

Bounding-box errors can impair control, and the finite skill vocabulary limits manipulation coverage. AR-Q trailed Cortex on RMBench's M(1) subset, 71.2% versus 75.2%. The paper also identifies difficulty with fine spatial details, auxiliary planning and data-limited skill precision.[1]

## References

1. Jiang, Yicheng, Gan, Zesen, et al. ["AR-WAM: A Visual-Conditioned Agent-Ready World Action Model for Robotic Manipulation."](https://arxiv.org/html/2609.23578v2) arXiv:2609.23578, September 20, 2026; version 2, October 2, 2026. Sections III-V and Tables I, IV-VI.
2. AR-WAM authors. ["AR-WAM" project page.](https://ar-wam.github.io/) Method overview and experimental tables. Accessed October 20, 2026.
3. AR-WAM authors. ["AR-WAM" reference implementation and README.](https://github.com/AR-WAM/AR-WAM) Method, installation, pretraining and policy-training instructions. Accessed October 20, 2026.
4. Jasonggg15. ["AR-WAM-RoboTwin-Dataset."](https://modelscope.cn/datasets/Jasonggg15/AR-WAM-RoboTwin-Dataset) ModelScope dataset card, updated October 8, 2026. Accessed October 20, 2026.
5. AR-WAM authors. ["stage1_robotwin_adaln.yaml."](https://github.com/AR-WAM/AR-WAM/blob/main/configs/stage1_robotwin_adaln.yaml) First-stage training configuration. Accessed October 20, 2026.
6. AR-WAM authors. ["stage2_robotwin_adaln.yaml."](https://github.com/AR-WAM/AR-WAM/blob/main/configs/stage2_robotwin_adaln.yaml) Second-stage training configuration. Accessed October 20, 2026.