SCDP: Learning Humanoid Locomotion from Partial Observations via Mixed-Observation Distillation

Milo Carroll; Tianhu Peng; Lingfan Bao; Chengxu Zhou; Zhibin Li

cs.ROMar 10, 2026

SCDP: Learning Humanoid Locomotion from Partial Observations via Mixed-Observation Distillation

Authors: Milo Carroll, Tianhu Peng, Lingfan Bao, Chengxu Zhou, Zhibin Li

Organizations: Department of Computer Science, University College London

Abstract

Distilling humanoid locomotion control from offline datasets into deployable policies remains a challenge, as existing methods rely on privileged full-body states that require complex and often unreliable state estimation. We present Sensor-Conditioned Diffusion Policies (SCDP) that enables humanoid locomotion using only onboard sensors, eliminating the need for explicit state estimation. SCDP decouples sensing from supervision through mixed-observation training: diffusion model conditions on sensor histories while being supervised to predict privileged future state-action trajectories, enforcing the model to infer the motion dynamics under partial observability. We further develop restricted denoising, context distribution alignment, and context-aware attention masking to encourage implicit state estimation within the model and to prevent train-deploy mismatch. We validate SCDP on velocity-commanded locomotion and motion reference tracking tasks. In simulation, SCDP achieves near-perfect success on velocity control (99-100%) and 93% tracking success in AMASS test set, performing comparable to privileged baselines while using only onboard sensors. Finally, we deploy the trained policy on a real G1 humanoid at 50 Hz, demonstrating robust real robot locomotion without external sensing or state estimation.

Figures & tables

Fig. 1 : Deployment of Sensor-Conditioned Diffusion Policies (SCDP) on Unitree G1, performing robust locomotion at 50 Hz using only proprioceptive sensors, without external motion capture or state estimation.

Fig. 2 : Sensor-Conditioned Diffusion Policies (SCDP) architecture and training framework. The state-action diffusion policy conditions on sensor observation history $\{o_{t}\}$ , past actions $\{a_{t}\}$ , and commands $c_{t}$ , while predicting future trajectories containing privileged states $\{s_{t}\}$ and actions. This mixed-observation formulation enables learning global dynamics from partial sensing.

Panel of Figure 3 — Fig. 3 : Waypoint navigation task. Sequential frames showing the SCDP controlling the robot to navigate succesfully to all five waypoints (green: current target, red: remaining targets).

Panel of Figure 4 — Fig. 4 : Fast forward walking behaviour. Sequential frames showing the policy executing forward locomotion.

Panel of Figure 5 — Fig. 5 : Comparison of walking right and walking left sequences.

Panel of Figure 6 — Fig. 6 : Velocity tracking performance. Comparison of realized velocities against commanded targets (black) for (a) linear velocity $v_{x}$ and (b) yaw rate $\dot{\psi}$ . SCDP (orange) exhibits smoother tracking with reduced oscillations compared to the BeyondMimic baseline (teal), with slight lag during direction changes.

Method	$v_{\text{pelvis}}$ feedback	Perturb	Navigation	Joystick
BC ${}_{\text{co}}$	$\checkmark$	70.1	43.8	70.0
BC ${}_{\text{co}}$	$\times$	5.0	4.0	4.0
BeyondMimic ${}_{\text{priv}}$	$\checkmark$	99.4	98.6	100.0
BeyondMimic ${}_{\text{co}}$	$\checkmark$	20.7	22.2	3.0
BeyondMimic ${}_{\text{mo}}$	$\checkmark$	99.4	94.4	94.0
BeyondMimic ${}_{\text{mo}}$	$\times$	92.0	5.6	93.0

TABLE I : Success/completion rates (%) across observation modes. Subscript – priv: privileged states; co: context observations only; mo: mixed observations. Bold: best per task.

Noise	0.0 m/s		0.5 m/s		1.0 m/s
Level	Push	No Push	Push	No Push	Push	No Push
0.0	98.2	32.5	98.2	24.7	95.8	18.7
0.06	99.3	78.4	99.0	79.4	98.9	74.1
0.12	87.1	81.6	99.5	98.0	98.8	95.0
0.18	97.5	82.4	99.4	99.2	98.8	99.6
0.24	88.2	70.0	99.1	99.6	99.2	99.3

TABLE II : Perturbation recovery success rates (%) under varying data collection strategies, Color intensity indicates performance (green=high, red=low).

Ours ${}_{\text{mo}}$
Context Length	$v_{\text{pelvis}}$ feedback	Perturb Succ. (%)	Nav. Compl. (%)
4	$\times$	84.3	74.2
8	$\times$	98.2	87.8
16	$\times$	55.5	33.4
SCDP
4	$\times$	99.5	97.8

TABLE III : Effect of context length on performance without velocity feedback. Bold indicates best per task.

		Train					Test
Method	$v_{\text{pelvis}}$ feedback	MPJPE-G $\downarrow$	MPJPE-L $\downarrow$	Vel-Dist $\downarrow$	Accel-Dist $\downarrow$	Success $\uparrow$	MPJPE-G $\downarrow$	MPJPE-L $\downarrow$	Vel-Dist $\downarrow$	Accel-Dist $\downarrow$	Success $\uparrow$
MMP (upper bound)	$\checkmark$	69.84 $\pm$ 25.96	47.73 $\pm$ 23.82	3.03 $\pm$ 1.07	1.00 $\pm$ 0.47	98.94	69.22 $\pm$ 26.12	46.58 $\pm$ 22.99	3.07 $\pm$ 1.04	1.03 $\pm$ 0.46	99.03
BC	$\times$	370.36 $\pm$ 278.93	53.65 $\pm$ 23.15	12.87 $\pm$ 7.59	14.78 $\pm$ 13.91	30.19	371.66 $\pm$ 276.42	50.71 $\pm$ 19.66	6.39 $\pm$ 3.01	1.31 $\pm$ 0.55	31.24
$\text{Ours}_{mo}$	$\times$	457.87 $\pm$ 374.15	54.81 $\pm$ 22.82	6.16 $\pm$ 2.33	1.33 $\pm$ 0.48	78.39	473.91 $\pm$ 408.60	54.36 $\pm$ 21.98	6.41 $\pm$ 2.52	1.39 $\pm$ 0.51	79.17
SCDP	$\times$	282.43 $\pm$ 167.54	53.63 $\pm$ 23.60	5.37 $\pm$ 2.14	1.48 $\pm$ 0.70	93.12	288.30 $\pm$ 184.22	53.12 $\pm$ 23.03	5.55 $\pm$ 2.25	1.51 $\pm$ 0.74	93.22

TABLE IV : Motion tracking performance on AMASS-walk dataset. MMP: expert upper bound.

Mixed Obs.	Representation	Context Dist.	Full Prefix	Restricted	Navigation
Training	Alignment	Alignment	Attention	Denoising	(%)
All features
$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	89.2
No features
$\times$	$\times$	$\times$	$\times$	$\times$	1.4
Section 1: Exclusion Ablation (all except one)

TABLE V : Ablation study on key design components for diffusion policies without velocity feedback.

Explore similar work

RoboDreamer: Anticipatory Humanoid Locomotion with Predictive State-Space Models
Sep 7, 2026Zhe Li, Yangyang Wei, Xichen Yuan +4Robot Trajectory Tracking Humanoid Robot Locomotion
BAAI · NJU
HuMBLE: Human Motion-Driven Behavior Learning for Embodied Locomotion
Oct 7, 2026Mike Zhang, Dongho Kang, Kevin Bergamin +19Humanoid Robot Locomotion RL for Legged Locomotion
RAI Institute, USA · Boston Dynamics, USA · Carnegie Mellon University, USA
ADAPT: Analytical Disturbance-Aware Policy Training for Humanoid Locomotion
Jun 15, 2026Bofan Lyu, Jindou Jia, Kuangji Zuo +7Robotic Control Humanoid Robot Locomotion
1MARS Lab, Nanyang Technological University

Sep 7, 2026J/K move · Enter open · S save

RoboDreamer: Anticipatory Humanoid Locomotion with Predictive State-Space Models

Zhe Li, Yangyang Wei, Xichen Yuan, Zhenzhe Zhang+3

BAAI · NJU

Robot Trajectory Tracking Humanoid Robot Locomotion Robust RL

Open paper AskPDF arXiv

Humanoid locomotion requires control policies that remain stable under imperfect sensing while exploiting temporal context for consistent motion. We present RoboDreamer, a two-stage teacher--student framework that combines next-observation consistency with randomized continuous temporal masking. A teacher is first trained on clean observations, and a student is then distilled under masked recent observations, encouraging the policy to infer missing current information from history. At inference, the same masking interface is reused for implicit closed-loop action refinement and optional multi-step action chunking. Mamba is used as the temporal backbone, while matched ablations show that masking/distillation provides a substantial part of the gain and Mamba contributes additional tracking improvements with real-time latency. Experiments in IsaacLab, MuJoCo, and on a Unitree G1 demonstrate robust motion tracking under observation masking and successful real-world deployment.

Abstract

Explore similar work

RoboDreamer: Anticipatory Humanoid Locomotion with Predictive State-Space Models

HuMBLE: Human Motion-Driven Behavior Learning for Embodied Locomotion

ADAPT: Analytical Disturbance-Aware Policy Training for Humanoid Locomotion