Pose targets, not velocities
Navigation predicts smooth relative pose increments, so errors do not accumulate as quickly over a long approach.
PROJECT RESOURCES
FUDAN UNIVERSITY / FYSICS AI
UniWAM.
Unified mobile manipulation.
Unified Mobile Manipulation via Mixed-Stream World-Action Modeling and Manipulation Anchor Pose Supervision.
MOBILE MANIPULATION / REAL ROBOT LIVE LOOP
01 APPROACH
02 CLEAR CUPS
RESEARCH FROM
Fudan University
Fudan UniversityAT A GLANCE
Designed for the full mobile-manipulation loop.
1.5M+MAP-Data episodes
7.5Khours of navigation
9.85 cmMAP-Bench position error
80.0%mobile-manipulation success
THE GAPMobile manipulation
Mobile manipulation
breaks at the handoff.
Navigation needs a wide, forward view and smooth base motion. Manipulation needs close-up cameras and high-frequency end-effector control. Treating both as one undifferentiated action stream wastes data and loses the moment where the robot should stop.
01
Different eyes
Forward navigation views and local manipulation views have different visual requirements.
02
Different actions
Base pose increments and end-effector trajectories live in different action spaces and control rates.
03
One missing target
Semantic goals say what to reach, but not the exact position and heading where manipulation should begin.
UniWAM keeps a shared world model and gives each stream its own interface.
THE APPROACHShared world model.
Shared world model.
Specialized control.
Each example trains one stream at a time. Navigation and manipulation are packed as separate batch rows, so the same video and action DiTs learn from both without cross-stream attention.
STREAM INTERFACE
Navigation stream
Find the pose
that enables the task.
RGB forward view + language + base state
Stream inputs + image-plane targets
RGB viewlanguagebase stateMAP / boxdx dy dtheta
Pose increments over a 32-step horizon Manipulation stream
See the contact.
Execute the intent.
Wrist + head views + language + proprioception
Stream-specific observations
head viewwrist viewslanguageproprioceptionxyz R6D gripper
Camera-frame end-effector trajectoriesSame backboneIndependent samplingBatch-parallel inference
01
Stream-specific inputs
Each behavior sees the cameras and state variables it actually needs.
02
Independent supervision
Navigation-only and manipulation-only data train the same model directly.
03
One deployable policy
Both streams run independently or in one batch during a real task.
THE DATA ENGINEStop where the
Stop where the
task starts.
Manipulation Anchor Pose (MAP) is an explicit base position and heading for the next interaction. The automated pipeline turns large indoor 3D scenes into dense, image-plane supervision for the entire approach.
MAP =Where to stop.
How to face.
A terminal pose defined by scene geometry, then projected into every observation frame.
- 300
- held-out targets
in MAP-Bench - 109
- indoor scenes
in evaluation - 15K
- evaluation
trajectories
REAL-ROBOT DEMOSEvery source clip.
Every source clip.
Classified by behavior.
All full-length real-robot videos from the supplied archive are grouped as pure manipulation, mobile manipulation, and pure navigation. Every clip is included below with an English, source-grounded title.
FIXED-BASE CONTROL
11 source clipsPure manipulation
IN FOCUS
Coffee preparation & interactive manipulation
BASE + ARM CONTROL
3 source clipsMobile manipulation
Reverse movements were teleoperated; all forward motions were driven by model inference.
Evaluated onAgileX Cobot MagicAgiBot G2Dual-arm Franka Research 3AgileX single arm
MAP-BENCH / PRIMARY EVALUATIONManipulation-ready navigation,
Manipulation-ready navigation,
measured.
On MAP-Bench, explicit MAP supervision and mixed-stream training reduce both terminal position and heading errors. UniWAM-Scaled reaches the lowest error in every distance range, including long-range approaches.
UNIWAM-SCALED / OVERALL
9.85 cm
MAP-Bench position error
30.1% lower than the strongest external baseline
UNIWAM-SCALED / OVERALL
2.30 °
MAP-Bench heading error
44.0% lower than the strongest external baseline
SECONDARY · REAL-ROBOT / 24 TASKSlower is better · success higher
| Model | MAP nav pos / head | Mobile manipulation success | Manipulation success |
|---|---|---|---|
| pi0.5 | 23.3 cm / 5.15° | 62.5% | 71.4% |
| GR00T N1.7 | 18.1 cm / 6.55° | 39.4% | 36.8% |
| UniWAM-Base | 15.1 cm / 4.91° | 63.1% | 44.3% |
| UniWAM-Scaled | 11.7 cm / 4.25° | 80.0% | 68.9% |
↳
On long-range MAP-Bench approaches, UniWAM-Scaled’s position error grows by 1.9 times from short to long range, compared with 6.4 times for pi0.5.
UNDER THE HOODDetails that make
Details that make
the streams travel.
Image-plane supervision
Target boxes and projected MAP points keep both streams grounded in what the robot sees right now.
Camera-frame end effectors
Similar visual motions map to similar actions across four embodiments, making mixed data reinforce one another.
OPEN SOURCE RELEASEBuild the next
Build the next
mobile manipulator.
Code, data, and benchmark artifacts will be released by Fudan University and Fysics AI.
CITE UNIWAM
UniWAM Technical Report: Unified Mobile Manipulation via Mixed-Stream World-Action Modeling and Manipulation Anchor Pose Supervision
Wei Xue, Keliang Liu, Mingzhang Cui, Jinhua Xie, Jinjie Wei, Jianan Hou, Jingcheng Lu, Lintao Wang, Kaixiang Qiu, Yizhou Liu, Xinghai Ye, Jinghang Han, Mingcheng Li, Jie Gu, Shunli Wang, Lihua Zhang, Dingkang Yang