TARC: Time-Adaptive Robotic Control

TL;DR

Fixed-frequency robotic control forces a trade-off between efficiency and robustness. Time-Adaptive Robotic Control (TARC) enables a policy to output both an action $a_t$ and an application duration $\Delta t$, creating a variable frequency $f = f_{max}/\Delta t$ within the TaCoS extended-MDP and is trained offline with PPO before zero-shot deployment. Across a drifting RC car and a Unitree Go1 quadruped, TARC matches or surpasses fixed-frequency baselines in task rewards while substantially reducing control frequency and exhibiting state-dependent adaptation under perturbations. This approach reduces computation, energy use, and wear while maintaining performance, enabling more robust and efficient real-time robotic control in diverse tasks.

Abstract

Fixed-frequency control in robotics imposes a trade-off between the efficiency of low-frequency control and the robustness of high-frequency control, a limitation not seen in adaptable biological systems. We address this with a reinforcement learning approach in which policies jointly select control actions and their application durations, enabling robots to autonomously modulate their control frequency in response to situational demands. We validate our method with zero-shot sim-to-real experiments on two distinct hardware platforms: a high-speed RC car and a quadrupedal robot. Our method matches or outperforms fixed-frequency baselines in terms of rewards while significantly reducing the control frequency and exhibiting adaptive frequency control under real-world conditions.

Figures (9)

• Figure 1: The performance of TARC on scenarios requiring adaptation in control frequency. The quadruped's control frequency spikes when experiencing a push, demonstrating state-dependent frequency modulation.
• Figure 2: Overview of our framework. The RL agent outputs an action $a_t$ and an application duration $\Delta t$, which allows the policy to modulate control frequency. We train completely offline in simulation (left), and deploy the policy zero-shot on the hardware (right).
• Figure 3: Desired reverse parking maneuver which involves rotating the car 180° and parking approx. 2m away
• Figure 4: The three evaluation scenarios for the Unitree Go1, defined by their command profiles over a 20-second episode. Each plot shows the commanded forward velocity (solid blue line, left axis) and yaw rate (dashed orange line, right axis) over time for each scenario (a) Gentle Curve, (b) Velocity Changes, and (c) Run Then Turn.
• Figure 5: Performance comparison of controllers on the RC Car task averaged over 5 seeds. The subplots show: (a) Total reward penalized with $c=0.1$. (b) Total reward excluding the penalty (c) Average control frequency over the episode

Theorems & Definitions (2)

• Definition 1: Time-Adaptive Policy $\pi$
• Definition 2: Reward function $R$