Qwen has developed three AI models that are embodied. They have been grouped together as Qwen Robot Suite. Qwen RobotManip is one of the three, followed by Qwen RobotWorld and Qwen RobotNav. Each robot is built using a Qwen-based vision language and is designed to solve a specific robotics issue.
Qwen-RobotManip is a Vision-Language-Action model for manipulation, built on Qwen3.5-4B. Qwen-RobotWorld has a Qwen2.5-3V encoder, 60 layer MMDiT with language conditioning and an embedded Qwen2.5-3V encoder. Qwen-RobotNav uses Qwen3VL and is available in sizes 2B to 8B.
Qwen-Robot-Suite
Qwen Robot Suite does not consist of a single robot model. The Qwen-Robot-Suite is made up of three different foundation models. RobotManip & RobotNav come with public GitHub repository.
Data on robots and robotics tasks is dispersed. The observation and actions formats used by different robots are incompatible. It is rare that a policy learned on one arm can be transferred to the other.
Three research reports have addressed this fragmentation differently. RobotManip aligns action representations so manipulation data scales. RobotWorld makes use of language to create a single action interface. RobotNav provides a configurable observation interface to perform navigational tasks.
This is the main split of the three releases:
| Model | The Problem | Backbone | Output |
|---|---|---|---|
| Qwen-RobotManip | Robotic manipulation | Qwen3.5-4B (Qwen-VL) | Continuous Robotic Actions |
| Qwen-RobotWorld | Modeling the world in a new way | Frozen Qwen2.5-VL | Video of the future: Prediction |
| Qwen-RobotNav | Mobile navigation | Qwen3-VL (2B/4B/8B) | The trajectory of a waypoint |
The Qwen Robot Manip Alignment System Unlocks the Scale to Perform Manipulation
Qwen-RobotManip is a Vision-Language-Action (VLA) foundation model. The model is based on Qwen VL, and it predicts robot continuous actions.
VLAs take both camera images and language instructions. This model outputs robot actions at a low level. Data manipulation is a heterogeneous problem.
Robots can record state and action in different formats. Scaling data can cause interference when demonstrations are delivered with incorrect representations. RobotManip’s unified framework for alignment solves the problem.
Unified Alignment Framework
Three complementary mechanisms are included in the framework. First It is a state-action canonical representation. It’s an 80-dimensional vectr with per-dimension masking.
Each block stores joint positions, end-effector pose, gripper state and dexterous hand joints. The block contains joint position, end effector pose, gripper status, dexterous hands joints, etc. Only the robots’ dimensions are filled.
The Second This is the camera frame delta pose parameterization. The deltas of the camera frame are what is used to represent end-effectors. The numerical proximity of motions across embodiments is visually the same.
Third It is a policy adaption mechanism that operates in context. It uses recent executions as an implied embodiment identifier. Policy adjusts behaviour at deployment without parameter updates.
This is accompanied by a co-training dual stream strategy. This optimizes both the manipulation of data as well as a stream for vision and language. It prevents backbone perception and reasoning from being eroded.
Data Engine
RobotManip has gathered approximately 38,100 hours worth of data manipulation. The only datasets used are open-source and videos of humans. It does not use proprietary data.
This scale is produced by a pipeline that converts egocentric hand demonstrations into robot trajectories. The pipeline converts hand-held demonstrations of egocentric movements into robot trajectory. This pipeline is rendered across 15 different robot platforms.
About 24,808 hours worth of demonstrations can be obtained from this synthesis. About 1,933 hours of data are derived from the egocentric sources. The open-source robot datasets account for over 11000 hours.
Visual alignment is separated from action alignment by the pipeline. The pipeline separates action alignment from visual alignment. Visual alignment using SAM3 Masking, ProPainter Inpainting and MuJoCo Inverse Kinematics.
It then filters out the corpus in a pipeline of five stages. The system detects sudden shifts, temporal errors, and extreme values. A check revealed that 81% of the episodes from a specific subset were not in alignment with state-action.
Benchmark Results
According to the research report, standard benchmarks do not measure generalization. In distribution tests, models without pre-training robots are comparable to those that have been pretrained. RobotManip is therefore focused on the settings used outside of distribution.
| Benchmark (OOD) | Prev. SOTA (π0.5) | Qwen-RobotManip |
|---|---|---|
| LIBERO-Plus | 84.4 | 91.4 |
| RoboTwin C2R Hard | 47.9 | 69.4 |
| EBench | 27.1 | 45.6 |
| RoboCasa365 | 16.9 | 35.9 |
| RoboTwin-IF | 49.6 | 72.2 |
Most of the reported differences are in cross-embodiment. RobotManip can reach 23.9% by using EEF camera frame actions. That is 3.2× the 7.5% achieved by π0.5.
This model is also ranked 1st in the RoboChallenge Table30 v1 Generalist Track. The model scores 20% better than the previous best. The AgileX Franka and ALOHA platforms are covered by the Real-Robot Validation.
Qwen RobotWorld: The Universal Action Interface, Language
Qwen-RobotWorld It is a video-world model conditioned by language. The model predicts the future of visual trajectory based on a recent observation. It uses natural language as the interface for unified actions.
The world model can learn the dynamics of an environment. It predicts the state of an environment given a state at present and a recent action. RobotWorld displays states and actions in video frames.
It is also important to note that language does not have a concept of embodiment. A single instruction encodes an action sequence, a goal and any constraints. This works with a Franka grasper, an Aloha Dual-Arm System, or even a Humanoid.
Double-Stream MMDiT Architecture
Model uses 60 layer Multimodal Transducer with Double Stream. A Qwen2.5 VL encoder is processed by an understanding stream. A generation stream processes video-VAE latents.
Two streams are interconnected by a joint focus at all layers. Two advantages come from using an MLLM to encode the actions. It can parse compositional directions and restrict physically plausible changes.
The MMDiT is 20B parameterized. VAE uses the Wan VAE architecture. Context length can support up to 48 360 video tokens.
This backbone is reused by a Scene2Robot system for cross-embodiment synthesizing. The system processes the scene, robot references, and generation segments simultaneously. It allows for human-to robot video transmission without the need to prompt robots.
Embodied World Knowledge Dataset
Embodied World Knowledge dataset (EWK). This dataset consists of approximately 8,6M video/text pairs. This is over 200 million observation frames.
This corpus includes four domains of embodied video plus general video. Manipulation provides about 5.9M samples across 20+ morphologies. The rest is made up of driving, navigation and transfer from human to robot.
A framework for mapping action to language standardises everything. It converts more than 500 categories of action and over 20 types of embodiment into language. Captions are generated by an hierarchical annotation pipeline consisting of five layers.
Benchmark Results
RobotWorld has been evaluated against four benchmarks. The game is the best on two out of four benchmarks.:
| Benchmark | Results | Ranking |
|---|---|---|
| EWMBench | 4.60 | Overall, 1st place |
| DreamGen Bench | 4.952 | Overall, 1st place |
| WorldModelBench | 8.99 | 1st open-source (3rd overall) |
| PBench | 0.804 | 1st open-source |
EWMBench has the highest HSD at 0.566. This is an increase of 33% over the second-placed. Scene consistency reaches 0.914.
WorldModelBench gives it a score of 1.00 in four categories. Newton’s law, mass conservation laws, fluid dynamics and gravity are the four categories. The following are the scores: Penetration 0.94 and Instruction Following 2.33.
Qwen-RobotNav : A controllable interface for navigation
Qwen-RobotNav This is a navigation model that can be scaled up using Qwen3 or VL. The model reframes navigation of multiple tasks as context-based observation. It exposes a parameterized control interface.
Navigation encompasses a wide range of tasks. Navigation is divided into many different categories. Each requires a distinct strategy to consume the visual stream.
For the purpose of instruction following, it is necessary to have a good memory in order to refer back and forth between landmarks. Only the latest frames are needed for target tracking. A fixed strategy for context is not the best way to accomplish all tasks.
Parameterized Interface
RobotNav formulates every task as a waypoint trajectory. The software predicts eight waypoints with each having a 2D location and heading. The lightweight MLP 4-layer head creates these.
There are two dimensions to the interface. Navigation behavior is selected by task mode for VLN (Visual Landing Network), PointNav (PointNav), ObjNav and Tracking. The visual history encoded by observing parameters is controlled.
The observation control includes a budget for visual tokens and a decaying time. Also included are weights for each camera. Randomization of all parameters during training time ensures robustness.
Natural-language tags are used to identify the camera and order of time. It requires no architectural changes to Qwen3VL. In order to support the new platform, it is only necessary to create a prompt template.
Agentic System
RobotNav is a great building block to build agentic systems. The upper-tier planner breaks long-horizon objectives down into sub-goals. The system’s planner for this level is Qwen3.6.
The planner reconfigures RobotNav’s Task Mode mid-episode. RobotNav acts as a reactive executor. Two tiers of communication are conducted exclusively in natural language.
Two-level memory is a support for long-horizon reasoning. Each rollout is summarized by a single-episode. Cross-episode memory accumulates durable conclusions like searched regions.
Benchmark Results
RobotNav has been trained using 15.6M sample data. This consists of 85% navigation trajectory data. Data on reasoning based on visual language fills out the rest of this 15%.
| Benchmark | Metric | Results |
|---|---|---|
| VLN-CE RxR (Val-Unseen) | Success Rate | 76.5% |
| VLN-CE R2R (Val-Unseen) | Success Rate | 72.1% |
| EVT-Bench | Tracking Rate | 90.0% |
| HM3Dv2 (ObjectNav) | Success Rate | 75.6% |
| NAVSIM | PDMS | 91.4 |
It is the new standard for Embodied Question-Answering. The system improves on the previous method in HM-EQA by 10,8%. The EXPRESS Bench also improved by 15%, while the navigation was 77% less.
This report demonstrates an improvement in performance from 2B up to 8B. The joint multi-tasking training creates a common spatial planning substrate. According to the report, this transfer occurs across different task families.
Use Cases With Examples
Each model maps to concrete deployment scenarios. Below are examples that combine report-supported findings with an illustrative frame.
- RobotManip: a few-shot deployment tool for new hardwareThey have a Franka hand and some demonstrations. RobotManip is fine-tuned on the workspace. This report shows that the pre-trained prior is more effective in reducing clutter and unnoticed states.
- RobotManip: cross-embodiment skills transferThe policy will be fine-tuned using 6K CobotMagic demonstrations and 130 ARX demos. The policy is tested with four new ARX tasks, but no target task demonstrations. The research reports 55.0% success, over 4× the best ablated variant.
- RobotWorld is a synthetic engine of dataThe VLA needs to collect more data for training than is possible through physical collection. As one of its three applications, the team of researchers lists generation synthetic data. RobotWorld generates video instructions for new languages.
- RobotWorld: a system for evaluating policyIn the research, policy evaluation is listed as a secondary application. It is possible to run a policy against generated paths before real hardware. It is a presentation of a general direction and not an actual benchmarked outcome.
- RobotNav in an Agentic System: An upper-tier planner decomposes a long-horizon goal into sub-goals. The system dispatches different navigation requests with different contexts and task settings. It improves by 10.8% over the previous EQA method.
- RobotNav for autonomous driving. This model also has a task for point-goal navigating. On NAVSIM, it reaches a PDMS of 91.4. By default, the forward camera is given the most weighted token.
Comparative Table of the Three Models
Below is a table that consolidates technical specifications. The table below is intended to be a guide for selecting the appropriate model.
| Attribute | RobotManip | RobotWorld | RobotNav |
|---|---|---|---|
| Work type | Manipulation of the VLA | Video World Model | You can navigate using |
| Backbone | Qwen3.5-4B | Frozen Qwen2.5-VL | Qwen3-VL |
| The Action Interface | Camera-frame EEF / joint | Natural language | The trajectory of a waypoint |
| Training data | 38,100 Hours | 8.6M video-text pairs | Samples of 15.6M |
| Key Architecture | The DiT headflow-matching system | 60-layer double-stream MMDiT | MLP action head |
| Headline results | RoboChallenge was the first to table30-v1 | DreamGen 1st place on EWMBench | 75% on RxR VLN |
| Output | Continuous Actions | The video that predicts the future | 8 waypoints (x, y, θ) |
| Public Repo | “Yes (GitHub)” | Blog only | “Yes (GitHub)” |
Three separate research reports don’t present one system. They cover complimentary layers when read in combination. RobotWorld is responsible for simulation, RobotManip, manipulation and RobotNav, mobility.
Canonical Action Vector
Understanding the RobotManip code is important. The mechanism allows robots of different models to use the same one. Here is an illustration that simplifies the idea of masking.
# Sketch of RobotManip for 80-dim vector.
2 x 29 diam per arm blocks plus 22 reserve dimensions = 80.
# The following is only an illustration, it does not represent the actual implementation.
CANONICAL_DIM = 80
The report identifies the following semantic groups:
ARM_GROUPS = {
"joints"Joint Positions:
"eef_pose"9. # 3D Position + 6-D Rotation
"gripper": 1, # parallel gripper width
"hand"Dexterous Hand Joints: 12
}
ARM_BLOCK = sum(ARM_GROUPS.values()) # 29
def build_masked_action(populated_groups, arms):
""Build an action vector with a binary mask per dimension.
Set of names for the groups that this robot will use.
arms: 1 for single-arm, 2 for dual-arm.
The rest of the dimensions are not visible.
"""
Action = [0.0] * CANONICAL_DIM
Mask [0] * CANONICAL_DIM
idx=0
_ range(arms).
Size of group in ARM_GROUPS.items():
Groups that are populated by a group:
If d is in the range (idx, size + idx), then:
Mask[d] Only here do gradients of 1# flow
Size = idx
If the arms are equal to 1:
Idx = BLOCK_ARM # Skip to second block
Return action mask
The gripper on one arm is a 7-DOF gripper.
_, mask = build_masked_action({"joints", "eef_pose", "gripper"}, arms=1)
print(sum(mask)) # -> 17 populated dims; the rest stay zero and masked
It is important to understand the concept of a binary mask per dimension. The gradients are only allowed to flow in the entries that have been semantically defined. It prevents erroneous supervision of degrees of freedom that are absent.
The loss due to flow matching is also masked by the same principle. The contribution of each sample is the same, regardless of whether there are multiple dimensions. It prevents robots with a higher number of slots dominating the optimization.
What you need to know
- Qwen has released three AI models that are embodied: RobotManip (also known as RobotWorld), RobotNav, and Qwen RobotSuite.
- RobotManip aligns data from robots into a single 80-dimensional action vector, and is ranked 1st in RoboChallenge table30-v1.
- RobotWorld is the only action-interface that uses natural language. It ranks first on EWMBench, DreamGen Bench and EWMBench.
- RobotNav introduces an interface with a customizable token-budget and achieves 76.5 % SR in VLN CE.
- RobotWorld, the third model, is only available as a paper. Two models come with public GitHub repositories.
Check out the Technical details The following are some examples of how to get started: PapersQwen-RobotManip, Qwen-RobotWorld” Qwen-RobotNav). Also, feel free to follow us on Twitter Don’t forget about our 150k+ML SubReddit Subscribe now our Newsletter. Wait! Are you using Telegram? now you can join us on telegram as well.
You can partner with us to promote your GitHub Repository OR Hugging Page OR New Product Launch OR Webinar, etc.? Connect with us

