Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash 92 tok/s
Gemini 2.5 Pro 59 tok/s Pro
GPT-5 Medium 22 tok/s
GPT-5 High 29 tok/s Pro
GPT-4o 94 tok/s
GPT OSS 120B 471 tok/s Pro
Kimi K2 212 tok/s Pro
2000 character limit reached

JAEGER: Dual-Level Humanoid Whole-Body Controller (2505.06584v2)

Published 10 May 2025 in cs.RO and cs.AI

Abstract: This paper presents JAEGER, a dual-level whole-body controller for humanoid robots that addresses the challenges of training a more robust and versatile policy. Unlike traditional single-controller approaches, JAEGER separates the control of the upper and lower bodies into two independent controllers, so that they can better focus on their distinct tasks. This separation alleviates the dimensionality curse and improves fault tolerance. JAEGER supports both root velocity tracking (coarse-grained control) and local joint angle tracking (fine-grained control), enabling versatile and stable movements. To train the controller, we utilize a human motion dataset (AMASS), retargeting human poses to humanoid poses through an efficient retargeting network, and employ a curriculum learning approach. This method performs supervised learning for initialization, followed by reinforcement learning for further exploration. We conduct our experiments on two humanoid platforms and demonstrate the superiority of our approach against state-of-the-art methods in both simulation and real environments.

List To Do Tasks Checklist Streamline Icon: https://streamlinehq.com

Collections

Sign up for free to add this paper to one or more collections.

Summary

An Expert Review of "JAEGER: Dual-Level Humanoid Whole-Body Controller"

The paper under review presents JAEGER, a dual-level whole-body controller purposed to enhance the robustness and versatility of humanoid robot control. The authors challenge the classical single-controller frameworks by proposing an innovative structure that decouples control of the upper and lower bodies into two independent controllers. This dual-level approach alleviates issues related to the curse of dimensionality and fault tolerance, ultimately supporting both root velocity (coarse-grained) and local joint angle (fine-grained) tracking.

Central to the paper is the incorporation of a MLP-based retargeting strategy that effectively translates human motion data to humanoid contexts. This involves leveraging the AMASS dataset and utilizing a curriculum learning approach, which initially applies supervised learning followed by reinforcement learning (RL). These methodologies facilitate the initialization of control policies and subsequent optimization towards a more optimal RL policy.

The experimental results are robust, demonstrating JAEGER's superiority over state-of-the-art methods like HumanPlus, Exbody, and OmniH2O across diverse metrics in both simulation and real-world environments. Notably, the JAEGER controller exhibited significantly lower mean absolute errors in tracking upper-body and lower-body joint angles, as well as in maintaining root velocities, compared to its competitors. For instance, in the root-based mode, angular velocity tracking errors using JAEGER were shown to deviate by as little as 5 degrees per second, showcasing notable improvement over the state-of-the-art.

The split-controller design provides notable advantages by limiting mutual interference between the upper and lower body control tasks, an inherent challenge observed in single-controller systems. This design aligns with the authors' primary hypothesis that addressing each body segment with tailored controllers increases the stability and effectiveness of humanoid locomotion and articulation tasks.

From a theoretical standpoint, the JAEGER system contributes to advancing the multi-agent system paradigm within robotics whole-body control—a functional reimagining of the humanoid robot's action space as a Decentralized Partially Observable Markov Decision Process (Dec-POMDP).

Practically, these results emphasize JAEGER's potential utility in realistic applications, where low-latency teleoperation and seamless adaptation of complex human-like motions are paramount. The ability to run at high control frequencies (above 1000 Hz) further attests to its applicability in high-demand settings such as surgical robotics.

While the research is promising, the authors acknowledge limitations such as the incapability of supporting keypoint position commands or executing highly complex poses outside the robot's mechanical or dynamic constraints. The substantial requirement for reward engineering also points to ongoing challenges associated with sim-to-real transfer.

In conclusion, JAEGER represents a significant contribution to the domain of autonomous humanoid control systems. This dual-level control mechanism not only shows strong numerical performance but also paves the way for further research into flexible, adaptable, and precise robot movement frameworks. Future developments could potentially explore expanding the controller's support for various posture-related commands and further minimizing the sim-to-real gap.

Ai Generate Text Spark Streamline Icon: https://streamlinehq.com

Paper Prompts

Sign up for free to create and run prompts on this paper using GPT-5.

Dice Question Streamline Icon: https://streamlinehq.com

Follow-up Questions

We haven't generated follow-up questions for this paper yet.