Learning Expressive and Compositional Motion Representation via Spectral Skills

Feiyang Wu; Chenxiao Gao; Chen Yang; Ye Zhao; Bo Dai; Anqi Wu

cs.ROSep 29, 2026

Learning Expressive and Compositional Motion Representation via Spectral Skills

Authors: Feiyang Wu, Chenxiao Gao, Chen Yang, Ye Zhao, Bo Dai, Anqi Wu

Organizations: Georgia Institute of Technology

Abstract

Robotic foundation models offer a promising path toward general-purpose humanoid robot control, often through hierarchical architectures. However, their effectiveness depends on the command interface between the planner and the controller, which must support accurate execution while remaining easy to predict, and ideally allow new behaviors to be composed from prior ones. In this work, we introduce spectral skills, a latent representation of this interface that meets these requirements through predictive representation learning. By design, spectral skills compactly encode short motion segments and are learned by predicting subsequent motion rather than reconstructing the encoder input. On a 29-DoF humanoid, a controller conditioned on spectral skills reduces global tracking error by 62% relative to the state of the art. The same frozen controller chains independently encoded skills without a separate transition policy. It also composes new behaviors by adding orthogonal directions to any compatible base skill, producing combinations unseen in the training data. We demonstrate tracking, chaining, and composition, as well as control through a language-conditioned planner, on Unitree G1 hardware. Project page: https://spectral-skill.github.io

Figures & tables

Figure 1: Skill chaining and composition with one controller. One frozen controller chains skills and adds or removes steering, executing a loop in physics simulation.

Figure 2: Overview. (a) An encoder $E_{\psi}$ maps a state history to a skill $z$ that linearly parameterizes a low-rank transition model. (b) One controller $\pi^{\mathrm{lo}}_{\theta}$ takes this skill from reference tracking, language-conditioned planning, or skill composition.

Figure 3: Real-robot deployment of the frozen controller on a Unitree G1. Top: tracking diverse whole-body motions. Middle: skill chaining, a side step, a $180^{\circ}$ turn and another side step, by switching skill codes. Bottom: knobs composed on one base skill. From left to right: base; right arm up; right arm up with increasing left turns (three panels) and increasing right turns (two panels) with each instance being a separate trajectory.

124-motion capability set
Method	SR $\uparrow$	MPJPE-L $\downarrow$	MPJPE-G $\downarrow$
Any2Track 1	69.4	$\approx$ 60	–
BeyondMimic 1	85.4	39.1	–
SONIC 1	99.2	23.8	–
SONIC	100.00	23.79	173.92
Ours	100.00	18.22	65.06

Table 1: Whole-body tracking. 1 Reported by SONIC on its own evaluation sets, not on these clips.

Direction	Measured effect	Direction	Measured effect
Left arm raise	shoulder pitch $\pm 0.86$ rad	Arms out	roll $+0.45$ rad/side, $+0.43$ (carry)
Both arms	shoulder pitch $\pm 0.67$ rad	High steps	swing apex $15\to 36$ cm (L), $14\to 32$ (R)
Right arm raise	shoulder pitch $\pm 0.90$ rad	Turn	$\pm 130^{\circ}$ in 4 s
Right arm yaw	shoulder yaw $\pm 0.43$ – $0.51$ rad †	Speed	$+0.33$ m/s, $6^{\circ}$ heading change ( $a{=}3$ )

Table 2: Effects on the walk base at

a{=}2

in Isaac PhysX.

Figure 4: One controller, many composed skills. Rows: base skills (walk, jog, squat, backward walk); columns: knobs added to the same frozen controller.

Figure 5: Steering composes skills rarely seen in the data. (a) Map of all $129{,}785$ training clips (t-SNE), colored by motion kind; (b) walking corner. (c)–(d) One walk steered along the arm, turn and high-steps directions: (c) steered skill on the arm and turn directions, with contours of corpus walks that raise an arm or turn; (d) distance of the robot’s motion to the corpus (dashed: unsteered) and corpus windows showing each combination.

Skill tracker: 256-D $z$ , hold 1
Interface	$N$	L $\downarrow$	G $\downarrow$	SR $\uparrow$
Oracle	—	17.19	51.7	1.000
Skills ( $z\!\to\!z$ )	30	38.41	251.2	0.914
Skills ( $z\!\to\!z$ )	10	42.83	230.2	0.911
Re-encode ( $o\!\to\!z$ )	10	60.19	425.1	0.771
Re-encode ( $o\!\to\!z$ )	21	88.64	796.8	0.546

Table 3: Language-conditioned planning interfaces. Arrows denote planner output

\to

tracker command;

z

is skill and

o

is explicit root-and-joint frame. L/G: local/global MPJPE (mm).

Appendix figures & tables20 assets

Supplementary material from the paper’s appendix.

Appendix

Setting	Value
Reference data	129,785 retargeted BONES-SEED clips
Robot and simulator	29-DoF G1; Isaac Lab with the Newton/MuJoCo-Warp backend
Control frequency	50 Hz
Offline pretraining	50,000 updates; batch size 8,192
Offline optimizer	AdamW; weight decay 0; gradient norm clip 1
Encoder learning rate	$3\times 10^{-4}$

Appendix

Table 4: Representation-learning settings.

Setting	Value
Tracker optimizer	AdamW; initial actor/critic LR $10^{-3}$
Actor LR adaptation	Per-iteration KL; target 0.02; limits $[10^{-5},10^{-3}]$
Ours: critic LR	Linear decay $10^{-3}\rightarrow 10^{-5}$ over the 50B budget
Ours: tracker weight decay	$10^{-2}$
PPO clipping; GAE; discount	$0.2$ ; $0.95$ ; $0.97$
Entropy coefficient; gradient norm clip	0; 1

Appendix

Table 5: Tracker-training settings.

Prediction target	SR $\uparrow$	MPJPE-L $\downarrow$	MPJPE-G $\downarrow$
Ours	92.85	24.47	99.51
End-point	91.48	26.24	128.32
End-point-det	91.92	26.25	110.77
Next-skill	92.46	25.32	102.42

Appendix

Table 6: Prediction-target ablation. All arms, including Ours , share one interface and are scored at the same

2.0

B-frame checkpoint on the same

4096

-motion set. Ours here is not the 50B tracker of Table 1 .

Representation	Training	SR $\uparrow$	MPJPE-L $\downarrow$	MPJPE-G $\downarrow$
Ours	offline, cont. 64	92.85	24.47	99.51
Ours	offline, cont. 256	92.77	24.68	104.02
Ours	offline, FSQ	89.28	31.19	108.87
Ours	offline, cat.	68.48	46.97	273.29
Recon.	offline, cont.	93.53	22.93	278.98
Recon.	offline, FSQ	93.68	22.50	238.12

Appendix

Table 7: Representation-learning ablation.

Design axis	Variant	SR $\uparrow$	MPJPE-L $\downarrow$	MPJPE-G $\downarrow$
Ours		92.85	24.47	99.51
Skill width	Cont. 128-D	92.90	24.79	99.81
Skill width	Cont. 256-D	92.77	24.68	104.02
FSQ	$64\times 32$	89.28	31.19	108.87
Codebook	Gumbel $64\times 32$	69.60	44.03	180.60
	Cat. $64\times 32$	68.48	46.97	273.29

Appendix

Table 8: Additional interface design choices.

Planner output	SR (%) $\uparrow$	MPJPE-L $\downarrow$	MPJPE-G $\downarrow$
Oracle	–	20.47	90.5
FSQ latent, temporal ensembling	90.4	45.49	220.5
FSQ latent, no ensembling	87.7	52.30	246.5
Explicit re-encode	90.2	44.75	252.0

Appendix

Table 9: Language-conditioned planning with an FSQ tracker (

64

-D, hold

10

). Every planner row re-plans every

N=10

control steps (one held FSQ skill, or

10

frames of the explicit chunk). Compare Table 3 .

Figure 6: Language-conditioned rollouts. The planner produces skills that the low-level policy executes: walking in a circle, lifting and carrying a box, opening a door while turning, and reaching from low to high.

Set	Method	SR	MPJPE-L	MPJPE-G	Root drift $d$	$\sqrt{d^{2}+\mathrm{L}^{2}}$	Final
124	SONIC	100.00	23.60	176.21	174.95	176.54	136
	Ours	100.00	18.00	65.05	58.66	61.35	15
4096	SONIC	98.88	26.52	190.82	189.27	191.12	177
	Ours	98.29	20.55	70.13	63.20	66.46	17
$4096^{\ast}$	SONIC	–	26.35	187.7	186.1	–	–
	Ours	–	20.48	70.0	63.1	–	–

Appendix

Table 10: World-frame error decomposition, re-measured with matched frame indices (one seed, one evaluation per method). MPJPE and root drift

d

are frame-weighted means over successful clips. Final is the median, over successful clips, of the root error at the last frame. All errors in mm. ∗ The

4{,}013

clips of the

4096

set that both methods complete.

Figure 7: Global root tracking, ours vs. SONIC. (a, b) Root position error over the first $8$ s of jointly completed clips lasting at least $8$ s: mean (line) and interquartile range (band). In (a), a few SONIC clips that drift by over a meter lift its mean above the upper quartile. (c) MPJPE-L, root drift and MPJPE-G over successful clips. The tick marks $\sqrt{\mathrm{drift}^{2}+\mathrm{L}^{2}}$ . (d) Mean root error per jointly completed clip of the $4096$ set. The dashed line is equality.

Figure 8: Four long clips from the $4096$ set, seen from above. Left: rollouts, one fixed camera per clip. Root paths: reference dashed, ours blue, SONIC orange. Robots are drawn at the numbered times, ours and SONIC shaded, the reference as an outline. The scale bar is $1$ m at the depth of the nearest path point. Right: root position error, with the pose times marked. In (b) and (c) the poses span the part of the clip in which the reference moves; (d) shows one pose, at the time of our largest error.

		Mean root error		Final root error
Clip	Reference motion	Ours	SONIC	Ours	SONIC
(a) Walk, random directions	$28.3$ s, $12.6$ m path	3.9	43.8	0.7	58.3
(b) Jog forward	$8.9$ m in $4.8$ s, then stands	2.7	62.1	1.6	76.2
(c) Walk sideways	$8.3$ m	2.2	55.4	1.3	89.5
(d) Balance on one foot	$24.3$ s	29.9	17.2	91.8	40.8

Appendix

Table 11: Root error of the four clips in Figure 8 , averaged over the clip and at the last frame, in cm.

Interface	Falls	Pose	Speed	Turn
Steering	0/66	1.00	0.95	$279^{\circ}$
Offset on action	17/66	1.30	0.78	$162^{\circ}$
Offset on servo target	5/66	0.76	0.91	$68^{\circ}$

Appendix

Table 12: Steering vs. matched joint offsets over

66

triples (5 bases, 5 directions, 3 amplitudes, and one wave gesture). Pose is the achieved pose relative to the steered run, so it is

1.00

for steering by definition; steered runs have

0

falls by construction. Speed is the median ratio to base for arm directions on walk, jog and backward walk. Turn is the turn direction’s median heading change. The speed ratios are not robust across seeds (Appendix G.1 ).

Method	Right arm	Spread	Speed	Turn
Ours, target-directed	9/12	9/10	4/12	6/11
Ours, spectral directions only	3/5	0/5	1/5	2/5
Joint offset (oracle)	2/10	8/9	0/10	5/10
SONIC, token PCA	4/5	2/5	0/5	0/5
SONIC, ridge directions	0/5	0/5	0/5	4/5

Appendix

Table 13: Steering pass rates at

a\geq 2

on five bases (walk, jog, squat, carry, backward walk). Target-directed directions are computed from chosen output rows of

J_{\xi}

; spectral directions are response-Gram eigenvectors. Joint offset (oracle) adds the matched steered run’s mean joint deviation to the servo target. Denominators count runs:

10

Isaac runs (

a{=}2

and

3

per base, one joint-offset spread run lost), plus

2

2

and

1

walk runs of our right-arm, speed and turn steering in the MuJoCo deployment simulator. Other rows have one run per base at

a{=}2

(SONIC in the MuJoCo simulator). Spectral-direction and token-PCA rows use directions chosen on walk; single seed.

Prompt	speed (m/s)	yaw ( ∘ )	R. shoulder pitch (rad)	other
Plain walk	0.60	22	$+0.21$
Raise right arm	0.69	81	$-1.20$	R. shoulder roll $-0.61$
Spread arms	0.64	49	$+0.29$	roll L $+0.55$ / R $-0.50$
Speed up	0.92	21	$+0.27$
Turn	0.61	88	$+0.22$

Appendix

Table 14: BFM-Zero reward prompts applied to a walk (hold window, single seed). Yaw is the heading change over the window.

Figure 9: A walk under our steering (top row) and under BFM-Zero’s reward prompts (middle row). Tiles show poses $1.2$ – $3.6$ s after onset. The bottom row shows ground paths from above (left turns bend upward). Numbers are changes against the same method’s plain walk over its hold window (ours $0.5$ – $5.9$ s, BFM-Zero $1.0$ – $6.0$ s after onset); plain-walk tiles instead give that walk’s own forward speed and heading change. Spread is the mean outward shoulder-roll change of both arms, and “speed $\times$ ” is the ratio of mean ground-path speeds. The “front” insets show the newest pose from the robot’s front right.

Line	$\lambda$ or $a$	Speed	Heading	R. shoulder	L. shoulder	L. foot apex
		(m/s)	( ∘ )	(rad)	(rad)	(cm)
Unsteered walk	–	$0.71$	$-2$	$-0.05$	$-0.04$	$15.6$
Blend toward standing, arm raised	$0.5$	$0.26$	$-50$	$-1.09$	$+0.10$	$8.0$
	$1$	$0.00$	$-11$	$-2.38$	$+0.27$	$0.0$
Arm-raise clip $-$ standing idle	$0.5$	$0.71$	$-42$	$-0.96$	$+0.17$	$20.6$
	$0.75$	$0.66$	$-55$	$-1.64$	$+0.33$	$23.7$

Appendix

Table 15: Raising the right arm during the walk (Isaac PhysX, walk base,

400

control steps, one seed; means over steps

65

–

335

). Blend:

(1-\lambda)z^{\rm walk}+\lambda z^{\rm arm}

; skill differences:

z^{\rm walk}+\lambda(z_{B}-z_{A})

; steering:

z^{\rm walk}+a\,v

, with

v

the right-arm spectral direction. Shoulder pitch is negative when the arm rises.

Stage	Speed (m/s)	Yaw ( ∘ /s)	Shoulder (rad)	Apex L/R (cm)	Distance	Support
Forward walk
Base	$0.98{\pm}0.03$	$-8{\pm}2$	$0.05{\pm}0.01$	$15{\pm}0$ / $16{\pm}0$	$2.73{\pm}0.09$ [2.76]	90009, 92259, 93794 [90252]
+ right arm	$1.07{\pm}0.01$	$3{\pm}1$	$-0.76{\pm}0.02$	$15{\pm}0$ / $16{\pm}0$	$3.32{\pm}0.04$ [3.37]	2334, 1988, 2124 [2180]
+ turn	$1.06{\pm}0.03$	$35{\pm}2$	$-0.77{\pm}0.03$	$14{\pm}1$ / $15{\pm}0$	$3.63{\pm}0.08$ [3.60]	112, 102, 100 [119]
+ clearance	$1.11{\pm}0.06$	$36{\pm}3$	$-0.80{\pm}0.04$	$19{\pm}2$ / $24{\pm}1$	$3.93{\pm}0.03$ [3.96]	19, 35, 95 [21]
Released	$1.02{\pm}0.01$	$-1{\pm}2$	$0.01{\pm}0.02$	$16{\pm}0$ / $15{\pm}1$	$2.87{\pm}0.15$ [2.69]	–

Appendix

Table 16: Composition ladder over three randomized seeds (mean

\pm

SD; brackets: deterministic run). Steering along the right-arm, turn and high-steps directions starts at

4

8

and

12

s (

a{=}2

1.2

and

1.2

), and all three are released together at

17

s. Each row averages the steady control steps of one stage. Speed is horizontal path speed, yaw the mean yaw rate, shoulder the mean right shoulder pitch, and apex the swing-foot height range over 1 s windows (left/right). Distance is the median over the stage of the Euclidean distance from the re-encoded executed skill to the nearest of

263{,}395

skills from

4{,}000

corpus clips (corpus skills to the nearest skill of another clip: median

2.80

, p90

4.01

). Support counts corpus 1 s windows (of

1{,}697{,}723

) that move like the base and show every attribute moved so far, one entry per seed. No run fell.

Interface	Falls	Pose	Speed	Drift ( ∘ )	Turn ( ∘ )
Steering	0, 0, 0 [0]	$1.00$	$0.95{\pm}0.02$ [0.90]	$16{\pm}3$ [20]	$295{\pm}10$ [279]
Offset on action	4, 5, 5 [5]	$1.37{\pm}0.03$ [1.38]	$0.98{\pm}0.10$ [0.89]	$36{\pm}8$ [29]	$192{\pm}4$ [177]
Offset on servo target	1, 1, 2 [1]	$0.73{\pm}0.00$ [0.76]	$0.92{\pm}0.01$ [0.91]	$23{\pm}3$ [19]	$92{\pm}13$ [80]

Appendix

Table 17: Steering vs. matched joint offsets at

a{=}2

over three randomized seeds (mean

\pm

SD; brackets: deterministic run). The

22

triples cover five bases

\times

five directions, minus three steered clearance runs that fell deterministically. Falls are per seed (

0

1

2

), out of

22

. Speed is the median ratio to base of arm directions on walk, jog and backward walk, drift the median

|

heading change

|

of non-turn directions, and turn the median heading change of the turn direction, all over runs that stayed up. Pose is relative to the seed’s steered run.

Setting	Frames	Success	MPJPE-L	MPJPE-G
Terminating, whole clip	6.0B	10/40	$22.5^{\dagger}$	$85.2^{\dagger}$
	27.0B	17/40	$22.1^{\dagger}$	$77.7^{\dagger}$
No termination, whole clip	6.0B	23/40 fall-free	110.9 (42.7)	829.9 (398.4)
	27.0B	24/40 fall-free	98.7 (27.8)	578.3 (142.0)
$20\times 10$ s windows per clip	6.0B	665/800 fall-free ‡	41.0 (31.0)	157.1 (127.6)
	6.0B	594/800 complete	–	–

Appendix

Table 18: Our method trained and evaluated on the 40 LAFAN1 clips. MPJPE in mm, mean over clips with the median in parentheses. No-termination rows include steps after a fall. Window rows average the pre-fall steps. † Successful clips only, frame-weighted. ‡ Torso-height fall test: torso below

0.4

m while the reference pelvis is at or above

0.4

m. With the relative fall test,

688/800

windows are fall-free at

27.0

Whole clip, terminating
	BFM-Zero	Ours, 27.0B
Success	0/40	17/40
Mean steps before termination	$1{,}028$	$7{,}904$
First failure: orientation / end effector / pelvis height	24 / 15 / 1	2 / 19 / 2
MPJPE-L / G, successful clips	–	22.1 / 77.7
Fall-and-get-up clips completed	0/6	0/6

Appendix

Table 19: BFM-Zero (released checkpoint, its MuJoCo deployment simulator and retarget) and our 27.0B LAFAN1 model (Isaac Lab) on the 40 LAFAN1 clips. MPJPE in mm, frame-weighted over the scored steps of the stated clips. The torso-height rows exclude every clip in which the torso reaches the floor: five fall-and-get-up clips for BFM-Zero, all six for ours. ∗ From a separate run in which a fall ends the clip.

Explore similar work

Jointly Learning Predicates and Actions Enables Zero-Shot Skill Composition
May 20, 2026Benedict Quartey, Sebastian Castro, Eric Rosen +3Visuomotor Policy Action Generation
Brown University · Robotics & AI Institute
Decompose and Recompose: Reasoning New Skills from Existing Abilities for Cross-Task Robotic Manipulation
May 2, 2026Xitie Zhang, Aming Wu, Yahong HanHuman-To-Robot Transfer Robotic Manipulation
School of Artificial Intelligence, College of Intelligence and Computing, Tianjin University, China · School of Computer Science and Information Engineering, Hefei University of Technology, China
SkillMemo: Expert-guided Skill Memory Framework for Compositional Embodied Manipulation
Aug 6, 2026Changyuan Wang, Chubin Zhang, Zhenyu Wu +8Visuomotor Control Real-World Manipulation Tasks
Shenzhen International Graduate School, Tsinghua University · Department of Automation, Tsinghua University · Nanyang Technological University +1

May 20, 2026, cs.ROJ/K move · Enter open · S save

Jointly Learning Predicates and Actions Enables Zero-Shot Skill Composition

Benedict Quartey, Sebastian Castro, Eric Rosen, Wil Thomason+2

Brown University · Robotics & AI Institute

Visuomotor Policy Action Generation Symbolic Predicates Generative Processes

Open paper AskPDF arXiv

Learning from Demonstration (LfD) enables robots to learn complex behaviors from expert examples, yet existing approaches often fail to generalize to new compositions of known skills without retraining. Modern generative policies model distributions over action trajectories alone, thus are unable to reason about the symbolic outcomes required for robust composition. We propose that skills should jointly model action trajectories and the symbolic outcomes they induce. To address this gap, we introduce Predicate Action Skills (PACTS), a class of closed-loop visuomotor policies that model skills as a joint generative process over action and predicate belief trajectories, producing coherent action-outcome rollouts within a single model. Jointly generating actions and predicates enables PACTS to learn internal representations that improve both action generation and predicate classification. Furthermore, we demonstrate zero-shot composition of learned skills via planning by leveraging online predicate predictions from PACTS as a symbolic interface for sequencing and monitoring execution. Project website: https://planpacts.github.io/

Abstract

Explore similar work

Jointly Learning Predicates and Actions Enables Zero-Shot Skill Composition

Decompose and Recompose: Reasoning New Skills from Existing Abilities for Cross-Task Robotic Manipulation

SkillMemo: Expert-guided Skill Memory Framework for Compositional Embodied Manipulation