# Hold My Beer: Learning Gentle Humanoid Locomotion and End-Effector Stabilization Control
**arXiv:** 2505.24198
**Authors:** Yitang Li, Yuanhang Zhang, Wenli Xiao, Chaoyi Pan, Haoyang Weng, Guanqi He, Tairan He, Guanya Shi
**Fetched:** 2026-02-13
**Type:** Research Paper

---

## Abstract

Can your humanoid walk up and hand you a full cup of beer, without spilling a drop? While humanoids are increasingly featured in flashy demos like dancing, delivering packages, traversing rough terrain, fine-grained control during locomotion remains a significant challenge. In particular, stabilizing a filled end-effector (EE) while walking is far from solved, due to a fundamental mismatch in task dynamics: locomotion demands slow-timescale, robust control, whereas EE stabilization requires rapid, high-precision corrections. To address this, we propose SoFTA, a Slow-Fast Two-Agent framework that decouples upper-body and lower-body control into separate agents operating at different frequencies and with distinct rewards. This temporal and objective separation mitigates policy interference and enables coordinated whole-body behavior. SoFTA executes upper-body actions at 100 Hz for precise EE control and lower-body actions at 50 Hz for robust gait. It reduces EE acceleration by 2-5x relative to baselines and performs much closer to human-level stability, enabling delicate tasks such as carrying nearly full cups, capturing steady video during locomotion, and disturbance rejection with EE stability.

## Key Contributions

- **Slow-Fast Two-Agent (SoFTA) framework:** Decouples upper-body and lower-body control into separate RL agents with different operating frequencies and distinct reward structures
- **Multi-frequency control:** Upper-body agent operates at 100 Hz for precise end-effector control; lower-body agent operates at 50 Hz for robust gait generation
- **Significant performance gains:** Achieves 2-5x reduction in end-effector acceleration compared to baseline whole-body control approaches
- **Near-human-level stability:** Performs much closer to human-level end-effector stability during locomotion
- **Practical task demonstrations:** Enables carrying nearly full cups without spilling, capturing steady video during walking, and maintaining EE stability under external disturbances
- **Addresses fundamental control mismatch:** Resolves the conflicting dynamics between slow-timescale locomotion and fast-timescale end-effector stabilization through temporal and objective separation

## G1 Relevance

SoFTA is **directly deployed and validated on the Unitree G1** humanoid robot (alongside the Booster T1). This makes it one of the most immediately applicable papers for G1 whole-body control. The dual-agent architecture with multi-frequency control is particularly relevant for G1 tasks that require simultaneous locomotion and manipulation — such as carrying objects, serving items, or maintaining stable tool use while walking. The framework comes from the same LeCAR Lab at CMU that produced H2O and OmniH2O, demonstrating a consistent research pipeline targeting the G1 platform.

## References

- Project Page: https://lecar-lab.github.io/SoFTA/
- GitHub: https://github.com/LeCAR-Lab/SoFTA
- arXiv: https://arxiv.org/abs/2505.24198