PaperlayerView exact revision on arXiv

Source-linked AI summary

PLAT: Sparse Timed Keyframe Motion Tracking for Humanoid Control via Privileged Latent Transition Learning

Zepeng Wang, Jiangxing Wang, Chao Ma, Xiaochuan Shi, Zongqing Lu

arXiv:2609.25754v1cs.RO

TL;DR

Sparse timed keyframe control asks humanoid policies to reach future poses at desired times without dense intermediate references. PLAT learns latent transitions from privileged dense sequences during training and refines them with residual reinforcement learning; it reports accurate, stable tracking across horizons and successful Unitree G1 deployment.

  • Problem

    Dense-reference tracking offers limited flexibility for applications that specify sparse goals, while timed keyframe control must infer stable transitions without intermediate references.

    Source: §I · p. 1

  • Method

    PLAT trains a dense tracking expert, distills privileged dense goal sequences into a sparse-command latent prior, then refines latent transitions with residual reinforcement learning.

    Source: §I · p. 1, B. Overview · p. 3, D. Privileged Latent Transition Learning · p. 4

  • Results

    PLAT reports accurate, stable sparse timed keyframe tracking across planning horizons, particularly long horizons, and successful transfer to a Unitree G1 humanoid.

    Source: §I · p. 2

  • Takeaways & Limitations

    Dense transition information can be used during training while the deployed controller operates from sparse timed keyframe commands.

    Source: §V · p. 8, §VI · p. 8

    PLAT relies on dense reference motions and a pretrained dense tracking expert during training, limiting applicability when privileged supervision is unavailable.

    Source: §V · p. 8

Abstract

from arXiv · show

Humanoid motion tracking policies rely on dense frame-by-frame references, limiting their use as high-level motion controllers for planning and interactive motion generation. We study \emph{Sparse Timed Keyframe Motion Tracking}, where a policy receives only sparse future keyframes and their desired arrival times, and must execute stable whole-body motions that reach successive goals. We propose \textbf{PLAT}, a three-stage sparse timed keyframe motion tracking policy learning framework with \textbf{P}rivileged \textbf{LA}tent \textbf{T}ransition learning. PLAT bridges dense motion tracking and sparse goal-conditioned control by exploiting dense goal sequences as privileged supervision during training while requiring only sparse timed keyframe commands at deployment. A pretrained dense tracking expert first provides robust motion priors. A privileged latent prior is then learned through DAgger-style imitation, followed by latent residual reinforcement learning that refines latent transitions instead of directly optimizing actions. Extensive simulation experiments demonstrate that PLAT maintains accurate and stable sparse timed keyframe tracking across varying planning horizons, with particularly strong performance under long-horizon commands. Successful deployment on a Unitree G1 humanoid robot further demonstrates the effectiveness and practicality of PLAT for sparse humanoid motion control.

I. INTRODUCTION

Dense frame-by-frame references make motion tracking effective for humanoid control but limit its use with sparse high-level commands. PLAT transfers dense motion knowledge into a policy that accepts timed keyframes at deployment.

A. Dense Motion Tracking for Humanoid Control

  • Dense motion tracking has enabled robust reproduction of locomotion, acrobatics, and whole-body behaviors, but requires a reference pose at every control step.Its local tracking formulation offers limited flexibility for interactive editing, planning, teleoperation, and language-driven motion generation.Source: §I · p. 1

I. INTRODUCTION

  • Sparse timed keyframe tracking asks a policy to reach future poses at commanded times while inferring stable transitions without dense references.Converting sparse commands into dense trajectories typically requires additional planning or motion generation modules, creating a multi-stage pipeline that is difficult to optimize jointly.Source: §I · p. 1
  • PLAT uses a pretrained dense tracking expert, privileged dense goal sequences for latent transition imitation, and latent residual reinforcement learning for refinement.The framework reports accurate, stable tracking across planning horizons, particularly long horizons, and successful transfer to a Unitree G1 humanoid.Source: §I · pp. 1–2

B. Sparse Motion Control

Sparse timed keyframe control requires inferring transitions that conventional sparse goal methods do not explicitly model. PLAT treats dense intermediate motion as privileged training information for a deployable latent transition prior.

B. Sparse Motion Control

  • Existing sparse-control approaches address goal reaching, motion completion, or behavior generation, while timed keyframe tracking also requires stable motion at commanded horizons.This introduces transition ambiguity absent from conventional dense tracking.Source: B. Sparse Motion Control · p. 2

C. Privileged Latent Representation Learning

  • PLAT's privileged information is the dense intermediate transition between sparse timed keyframes, rather than future robot observations or environment states.The resulting latent prior explicitly models motion evolution under sparse timed commands.Source: C. Privileged Latent Representation… · p. 2

A. Problem Formulation

  • The policy receives a sparse future keyframe and desired arrival horizon, then must infer physically plausible, temporally consistent whole-body motion without intermediate reference frames.The mapping from sparse goals to actions is inherently multimodal, and long-horizon transitions are challenging to infer from sparse commands.Source: A. Problem Formulation · pp. 2–3

B. Overview

PLAT progressively transfers dense motion knowledge into a sparse-command policy through three learning phases. At inference, only the robot state, a future keyframe, and its arrival time are required.

B. Overview

  • The three phases train a dense tracking expert, learn a sparse policy through privileged latent transition learning, and refine it with latent residual reinforcement learning.Dense goal sequences are available only during training; deployment does not require dense references or privileged transition information.Source: B. Overview · p. 3

C. Dense Motion Prior Learning

  • A fixed dense motion tracking expert provides stable motion priors and online action supervision for subsequent sparse-policy learning.It is trained with frame-by-frame references and guides the sparse policy through online imitation rather than being deployed directly for sparse keyframe tracking.Source: C. Dense Motion Prior Learning · pp. 3–4

D. Privileged Latent Transition Learning

A sparse destination leaves the intermediate motion ambiguous, especially as the planning horizon grows. PLAT uses dense transition sequences during training to teach a latent prior that can be inferred from sparse observations at deployment.

D. Privileged Latent Transition Learning

1) Privileged Transition Representation:

  • A GRU encodes the privileged goal sequence, and the prior predicts a latent transition distribution from deployable observations while the training posterior also uses the transition representation.The policy then predicts the expert action using the sampled posterior latent during training.Source: 1) Privileged Transition Representation · p. 4

2) Online DAgger Optimization:

Online DAgger supervises the sparse policy with expert actions on states induced by the evolving student, while transferring privileged transition knowledge to the deployment-time prior.

2) Online DAgger Optimization:

  • Online DAgger collects expert action labels on the evolving student policy’s state distribution, reducing compounding errors during long-horizon rollouts.The student interacts with the environment under sparse timed commands while the dense expert provides online supervision.Source: 2) Online DAgger Optimization · p. 4
  • Expert actions supervise both posterior and prior policies, and a KL divergence transfers privileged transition knowledge to the deployment-time prior.The supplied passages introduce the overall training objective but do not provide its full expression here.Source: 2) Online DAgger Optimization · p. 4
  • Deployment discards the posterior network and goal-sequence encoder, retaining the observation encoder, learned prior, and actor for sparse timed keyframe tracking.The deployed policy uses the current robot state, a target keyframe, and its desired arrival time.Source: 2) Online DAgger Optimization · p. 4

E. Latent Residual Policy Refinement

PLAT refines sparse tracking in latent transition space rather than directly optimizing actions, using a lightweight residual while preserving its learned motion prior.

E. Latent Residual Policy Refinement

  • Fine-tuning all policy parameters with PPO often degrades the learned motion prior and causes unstable behavior, so PLAT freezes the core networks and trains only a lightweight latent residual.The frozen components are the observation encoder, prior network, and policy decoder.Source: E. Latent Residual Policy Refinement · p. 5

1) Latent Residual Policy:

  • During reinforcement learning, the posterior network and privileged transition encoder are discarded, and the policy uses the deterministic prior latent as at deployment.The residual network predicts a correction to that latent.Source: 1) Latent Residual Policy · p. 5
  • Unlike action-level residual RL, PLAT adapts in latent transition space, preserving the learned transition prior while allowing task-specific optimization.The residual magnitude is controlled by α.Source: 1) Latent Residual Policy · p. 5

2) Reinforcement Learning Objective:

The latent residual policy is trained with rewards for dense motion tracking and sparse timed keyframe tracking, alongside smoothness and stability regularization. The experiments are framed around comparisons, horizon robustness, component contributions, and motion-transition quality.

2) Reinforcement Learning Objective:

  • The reinforcement-learning reward combines dense tracking, sparse timed keyframe objectives, and regularization for smooth, dynamically stable behavior.Dense tracking rewards preserve natural whole-body motion, while sparse rewards explicitly optimize timed keyframe tracking.Source: 2) Reinforcement Learning Objective · p. 5
  • After reinforcement learning, the deployed controller retains the observation encoder, latent transition prior, residual module, and policy decoder, using only robot state and a timed target keyframe.It requires neither dense reference motions nor privileged information.Source: 2) Reinforcement Learning Objective · p. 5

IV. EXPERIMENTS

  • The experiments ask whether PLAT improves tracking over dense policies, remains accurate and stable across horizons, and benefits from its latent-learning and residual components.They also examine whether privileged latent transition learning produces more natural and coherent transitions under sparse commands.Source: §IV · p. 5

A. Experimental Setup

Evaluation repeatedly issues future keyframes with normalized arrival horizons and measures performance near each commanded arrival time. The protocol includes single-step and mixed horizons, with three task-specific arrival-window metrics.

1) Evaluation Protocol:

  • Each episode repeatedly samples a future keyframe and normalized arrival horizon from a reference motion, issuing a new target when the horizon expires.This creates a continuous sequence of sparse timed commands throughout the motion.Source: 1) Evaluation Protocol · p. 5
  • The five evaluation settings test generalization across temporal horizons; Single uses Δ=1, while Mixed samples Δ uniformly from [1, 50] and is the default for quantitative results.Single degenerates to conventional one-step motion tracking.Source: 1) Evaluation Protocol · p. 5

2) Evaluation Metrics:

  • Metrics are computed only in the final three control steps before the commanded horizon expires, rather than at every simulation step.This arrival-window protocol is intended to reflect sparse timed keyframe tracking.Source: 2) Evaluation Metrics · p. 5
  • TKS@3 measures successful target reaches within the arrival window and termination thresholds; Timed MPJPE@3 and Timed MPKPE@3 measure mean joint and keypoint position errors there.All three metrics evaluate performance exclusively around the commanded arrival time.Source: 2) Evaluation Metrics · p. 5

3) Baselines:

The evaluation contrasts PLAT with dense-tracking baselines and tests how privileged latent transitions and residual refinement contribute to sparse timed keyframe tracking.

Baselines

  • SONIC and TWIST2 are dense frame-by-frame tracking policies used as baselines against sparse timed keyframe tracking.Both are designed for dense reference motions.Source: 3) Baselines · p. 5
  • Scratch RL trains a general tracking policy with PPO from scratch and also supplies the dense motion prior for later stages.Source: 3) Baselines · p. 5

Ablation variants

  • The w/o Privileged Latent variant removes the transition model, while DAgger is evaluated before reinforcement-learning refinement.Source: 3) Baselines · p. 6
  • Refinement variants compare full-policy PPO fine-tuning, action-space residual learning, and PLAT's latent residual refinement.Source: 3) Baselines · p. 6

B. Main Comparison

PLAT outperforms representative baselines under mixed horizons and maintains tracking performance across planning horizons. Ablations assess the latent transition model and latent residual refinement.

B. Main Comparison

  • PLAT achieves the highest TKS@3 and lowest Timed MPJPE@3 among the compared methods under the Mixed horizon setting.TWIST2 has the lowest Timed MPKPE@3 but substantially lower TKS@3, indicating frequent tracking failures.Source: B. Main Comparison · p. 6

C. Horizon Generalization

  • PLAT maintains high success rates and low timed pose errors across horizons, while Scratch RL degrades as keyframes become farther apart.SONIC and TWIST2 generalize poorly, with TWIST2 showing inconsistent pose errors across horizons.Source: C. Horizon Generalization · p. 6

D. Ablation Study

  • Adding a privileged latent transition model improves TKS@3 and Timed MPKPE@3 over deterministic DAgger from sparse observations.The ablation is reported under the Mixed horizon setting.Source: D. Ablation Study · p. 7, Table II · p. 7
  • Full-policy PPO fine-tuning collapses performance, action residuals improve stability but remain inferior, and latent residual refinement yields the strongest results.PLAT achieves the highest TKS@3 and lowest timed pose errors among the refinement variants.Source: D. Ablation Study · p. 7

E. Qualitative and Real-World Visualization

Qualitative comparisons show PLAT connecting sparse keyframes more smoothly than the dense-tracking comparison in simulation, with stable whole-body motion also demonstrated on a Unitree G1. The discussion notes training-supervision and horizon limits, while the conclusion describes accurate and stable sparse tracking in simulation and on hardware.

E. Qualitative and Real-World Visualization

  • In simulation, PLAT produces smoother, more physically plausible transitions and reaches target poses closer to commanded times than SONIC, while avoiding its abrupt or unstable recoveries.The paper also reports stable whole-body motion when PLAT is deployed on a Unitree G1.Source: E. Qualitative and Real-World… · p. 7

V. DISCUSSION AND LIMITATIONS

  • PLAT relies on dense reference motions and a pretrained dense tracking expert during training, limiting applicability when privileged supervision is unavailable.Very long planning horizons remain under-specified, making accurate transition inference increasingly difficult.Source: §V · p. 8

VI. CONCLUSION

  • PLAT transfers dense transition information into a sparse-command prior through DAgger-learned goal-sequence embeddings, then refines transitions with latent residual reinforcement learning.The conclusion reports accurate and stable tracking in simulation and on a Unitree G1 humanoid.Source: §VI · p. 8
Loading 2609.25754v1…