Humanoid controllers often require task-specific training despite sharing the need for balance and whole-body coordination. We present Metatron, a promptable whole-body execution layer combining a generative base policy with a programmable skill layer. The base policy receives partial goals, collectively called a prompt, that can combine timed body-part targets, velocity commands, and style cues. It produces diverse whole-body motions consistent with these goals. Training combines motion capture, synthetic recovery trajectories, and retargeted scene-interaction motions. After training, the policy is frozen. Stateful programs called skills express task intent as prompts at each control step, drawing rough geometric sketches that the policy turns into coordinated movement. Skills are ordinary code, refined by coding agents that analyze motion data and fit compact rules without task-specific policy training. On the physical Unitree G1, one frozen checkpoint supports walking, recovery, teleoperated object pickup, and sitting on furniture. Comparative evaluations show higher reaching success and push robustness than the tested reference-tracking pipelines. The same recipe produces climbing and acrobatics on a simulated character. Metatron enables coding agents to turn learned motor capabilities into new skills through code, without further policy training.
Figures & tables
Fig. 1 : One frozen controller per embodiment, diverse behaviors. Partial prompts specify goals; each controller provides whole-body coordination across tasks without task-specific training. Teal and gray blocks indicate specified and unspecified goal components, respectively.
Fig. 2 : Model and training. The prior maps partial prompts and observation history to a deterministic condition and a latent distribution. The decoder combines the condition with a sphere-projected latent to produce joint targets. Training adds a posterior conditioned on the full reference and supervision from a frozen tracker. Only the solid path runs at deployment. Pose renders illustrate state, not camera inputs.
Fig. 3 : Agent-assisted skill development. Given a task specification and optional motion examples, a human and LLM agent define evaluation scenarios and criteria. The LLM agent then iteratively refines the skill code using simulation feedback. Skills specify goals for the frozen Metatron policy, enabling the model to perform new tasks without additional training.
Fig. 4 : Change the prompt, retain the controller. In this simulation skill, a detected fall switches steering goals to a whole-body upright pose; steering resumes once upright. G1 deployment instead requests torso height alone.
Fig. 5 : Multiple control modes through one G1 policy. Each mode specifies different goals for the same frozen controller.
Fig. 6 : SOMA: On a simulated SOMA character, with the same recipe, Metatron runs, slides off a ledge, recovers from the fall, then runs and climbs up a wall.
Fig. 7 : Recovery: Metatron recovers from a range of initial configurations.
Method
Reached (%)
Error (cm)
Upright (%)
Metatron
83
5.9
100
ARDY + SONIC
39
6.4
82
GRU + IK + SONIC
28
5.6
81
GRU + SONIC
0
—
87
TABLE I : Whole-body wrist reaching on a 54-target grid. Metatron is trained without synthetic recovery data for this comparison. A target is reached if the wrist is held within 10 cm of it while the robot remains upright. Error is measured only on reached targets .
Perturbation force (N)
Method
200
400
600
800
1000
Metatron
100
100
62
75
75
ARDY+SONIC
100
100
17
0
0
GRU+SONIC
100
75
25
0
0
TABLE II : Push robustness during steering (100 ms force pulse at torso, 8 directions). Success: upright after 8 s. Metatron trained without recovery data. Results: success (%).
Perturbation force (N)
Method
0
200
400
600
800
Metatron
97.4−1.8+1.1
94.8−2.3+1.6
33.0−4.0+4.2
7.4−2.0+2.6
6.2−1.8+2.5
Global only
87.6−3.2+2.6
87.8−3.2+2.6
38.2−4.2+4.3
9.6−2.3+2.9
6.2−1.8+2.5
Mocap only
92.0−2.7+2.1
84.8−3.4+2.9
31.8−3.9+4.2
6.4−1.8+2.5
5.0−1.6+2.3
No aux.
81.6−3.6+3.2
78.8−3.8+3.4
29.2−3.8+4.1
5.4−1.7+2.3
4.6−1.5+2.2
MaskedMimic +
98.2−1.6+0.9
90.0−2.9+2.3
27.8−3.7+4.1
4.2−1.4+2.1
4.6−1.5+2.2
TABLE III : Push robustness during XY waypoint reaching. Horizontal torso pushes last 100 ms, recurring every 1–2 s during the first 5 s until arrival. Success: reaching within 0.5 m of the goal within 10 s, with relative torso height ≥0.45 m and tilt ≤60∘ at arrival. Entries report success (%) with upper/lower 95% Wilson CI deviations in percentage points over 500 episodes. Bold marks CI overlap with the highest rate at each force (unrounded bounds).
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Setting
G1 BONES
SOMA
Environments per GPU
8192
4096
Minibatch size per GPU
16384
8192
Optimization passes per rollout
2
1
Gradient-norm limit
1
50
Entropy coefficient
0.01
0.005
Reward normalization
No
Yes
Appendix
TABLE IV: Expert PPO settings, distinct from student distillation.
Term
w
σ
Torso, hands, feet: global position
1
0.3
Torso, hands, feet: global rotation
1
0.4
Torso and hands: global position
2
0.1
Torso and hands: global rotation
0.4
0.1
Anchor-relative body position
1
0.3
Anchor-relative body rotation
1
0.4
Appendix
TABLE V: G1 BONES expert rewards. Tracking terms are wexp(−e/σ2) . Position and velocity errors sum squared coordinate errors, then average over the selected bodies; rotation error is the mean squared relative rotation angle.
Skill
Input
Prompt fields
Update
Dense tracking
reference clip
timed body positions and orientations
reference clock
Steering
joystick velocity and turn
planar velocity; five timed torso, heading, and alternating foot targets
every step
Path following
waypoint list
timed torso waypoints and planar velocity
receding horizon
Location
target position or orientation
timed body constraint; planar position, height, heading, or full anchor orientation
on target change
Style
example clip
unordered motion-state cue
on reset or override
Sparse goal
pose from a trajectory
one timed future pose
on target transition
Appendix
TABLE VI : Harness skills and the prompts they construct. Each embodiment uses one frozen policy across its supported skills. The recovery row describes the simulation state machine; hardware get-up demonstrations are reported in section VI-B .
College of Engineering, Ocean University of China · Faculty of Information Science and Engineering, Ocean University of China · School of Artificial Intelligence, University of Chinese Academy of Sciences +1