Mobile humanoid manipulation requires both reaching a usable workspace and preserving precise hand-object interactions as object poses and contact conditions change. Learning these behaviors from limited task-specific data remains challenging. To bridge this gap, we introduce Praxis, a whole-body manipulation framework that combines physical interaction priors from one-shot egocentric video demonstrations with closed-loop posture calibration and online perception. The framework coordinates three stages: vision-language-guided navigation toward target objects, closed-loop posture calibration to align the arm-hand workspace, and dexterous manipulation with synchronized upper- and lower-body control. Online visual feedback re-grounds demonstrated interaction geometry under new object poses and scene configurations, while tactile feedback adapts hand motions to actual contact conditions. Each manipulation skill is specified by one human demonstration, without task-specific manipulation-policy retraining. Experiments across five long-horizon manipulation tasks demonstrate spatial, visual, and cross-object generalization, as well as recovery from external physical disturbances across all three stages.
Figures & tables
Fig. 2: Praxis architecture. Stage 1: Navigation to the manipulation region. Stage 2: Calibration to a manipulation-ready posture in a closed-loop manner. Stage 3: Distilling hand-object interaction priors through egocentric videos and applying online feedback to adapt manipulation policies.
Fig. 3: Hardware system of Praxis.
Fig. 4: End-effector IK error for the left and right arms with a fixed torso.
Fig. 5: Visualization of wrist pose detection and dexterous retargeting for the Pull Drawer task.
Hand Over
Stage 1
Stage 2
Stage 3: Manip.
Avg. Len.
Methods
Distance
Nav.
Calib.
pick cup
close to left hand
hand over
move to tray
place cup
Praxis
3 m (M.)
10/10
9/10
9/10
9/10
9/10
9/10
9/10
6.4
Praxis w/o Adapt.
3 m (M.)
1/10
1/10
1/10
0/10
0/10
0/10
0/10
0.3
Praxis
5 m (M.)
7/10
7/10
7/10
7/10
6/10
6/10
6/10
4.6
Praxis
3 m (M.+T.)
7/10
6/10
5/10
6/10
5/10
5/10
5/10
3.9
Praxis w/o Adapt.
3 m (M.+T.)
0/10
0/10
0/10
0/10
0/10
0/10
0/10
0.0
TABLE I: Step-wise success rates (successes/trials) and average number of completed steps (Avg. Len.) across navigation (Nav.), posture calibration (Calib.), and manipulation (Manip.). M.: translation-only approach; M.+T.: turning required. w/o Adapt.: original VLN policy. Colors denote the left ( blue ), right ( orange ), and both arms ( purple ).
Methods
Hand Over
Pour Water
pick cup
close to left hand
hand over
move to tray
place cup
Avg. Len.
pick cup
pick bottle
close to each other
pour water
place bottle
place cup
Avg. Len.
π0.5
29/33
26/33
18/33
16/33
15/33
3.15
26/33
28/33
25/33
20/33
18/33
17/33
4.06
YOTO
30/33
29/33
27/33
27/33
27/33
4.24
32/33
30/33
30/33
29/33
29/33
29/33
5.42
OKAMI
25/33
25/33
17/33
15/33
15/33
2.94
24/33
26/33
24/33
20/33
18/33
18/33
3.94
Praxis w/o Tact.
28/33
28/33
26/33
26/33
26/33
4.06
29/33
31/33
28/33
27/33
27/33
27/33
5.12
Praxis (Ours)
31/33
31/33
31/33
30/33
30/33
4.64
32/33
31/33
30/33
30/33
30/33
30/33
5.55
TABLE II: Quantitative performance comparisons on five long-horizon whole-body manipulation tasks. We report step-wise success rates and the average number of completed steps (Avg. Len.). Bold entries denote the highest Avg. Len. The two press–release cycles in Manipulate Pipette correspond to aspiration and dispensing.
Fig. 6: Visualization of varied object poses, scene configurations, and object categories, and failure recovery across three stages.
Step
Interference at
Stage 1&2
Stage 2
Stage 3
Stage 1&2
8/10
9/10
10/10
Stage 3: Manip.
pick cup
7/10
9/10
10/10
pick bottle
8/10
9/10
10/10
close to each other
7/10
9/10
10/10
TABLE III: Step-wise success rates and average number of completed steps (Avg. Len.) on recovery cases across all stages for Pour Water .
Step
Method
Praxis w/o Calib.
Praxis (Ours)
Stage 1&2
2/10
8/10
Stage 3: Manip.
pick cup
1/10
8/10
pick bottle
2/10
8/10
close to each other
1/10
8/10
TABLE IV: Step-wise success rates and average number of completed steps (Avg. Len.) under lateral interference across all three stages for Pour Water . Stage 1 Avg. Step is the average number of inference steps during navigation. Calib. denotes posture calibration.
Task
π0.5
YOTO
OKAMI
Praxis w/o Tact.
Praxis (Ours)
Hand Over
45.45
81.82
45.45
78.79
90.91
Pour Water
51.52
87.88
54.55
81.82
90.91
Pull Drawer
33.33
66.67
51.52
66.67
78.79
Open Lid
24.24
51.52
30.30
48.48
60.61
Manipulate Pipette
0.00
45.45
6.06
42.42
63.64
Mean
30.91
66.67
37.58
63.64
76.97
TABLE V: Success rates (%) across five tasks, with 33 trials per task. Mean denotes the average success rate across the five tasks.
Fig. 7: Cross-object reuse from one demonstration per task. Top and middle: Hand Over transfers a paper-cup demonstration to a toy bin and a plastic cup, respectively. Bottom: Pour Water transfers a bottle-and-paper-cup demonstration to a transparent bottle and a plastic cup.
Method
Dexterity
One-shot
Learn from Video
Vis. & Lang. Input
WholeBodyVLA [ 38 ]
✗
✗
✓
✓
HMC [ 39 ]
✗
✗
✗
✗
DemoHLM [ 40 ]
✗
✓
✗
✗
TrajBooster [ 5 ]
✓
✗
✗
✓
OKAMI [ 22 ]
✓
✓
✓
✗
Praxis (Ours)
✓
✓
✓
✓
TABLE VI: Comparison between recent whole-body manipulation systems and Praxis. Vis. & Lang. Input indicates vision–language input, and Long Dist. refers to long-distance navigation.
Action Type
Count
Move Forward
46
Turn Right
32
Turn Left
32
Move Right
20
Move Left
20
Total
150
TABLE VII: Distribution of trajectories in the fine-tuning dataset.
Fig. 8: Sequential snapshots of experiments in simulation. The top and bottom rows for each task display teleoperation-collected keyframes and their corresponding execution sequences, respectively. These results demonstrate the efficacy of our keyframe-based representation in maintaining precise interaction geometry across diverse manipulation tasks.
Fig. 9: Sequential snapshots showing the progression of the experiment across five tasks. Semi-transparent G1 indicates historical states.
Fig. 10: Visualization of wrist pose detection and dexterous retargeting in task Hand Over .
Fig. 11: Visualization of wrist pose detection and dexterous retargeting in task Pour Water .
Fig. 12: Visualization of wrist pose detection and dexterous retargeting in task Open Lid .
Fig. 13: Visualization of wrist pose detection and dexterous retargeting in task Manipulate Pipette .
Fig. 14: Visualization of fingertip tactile forces when grasping the toy (step pick toy ) in Open Lid using two-, three-, and five-finger configurations of the left Revo2 hand.