Bridging the Sim-to-Real Gap with Dynamic Compliance Tuning for Industrial Insertion

TL;DR

Experimental results show that the proposed framework, trained in simulation on a generic square peg-and-hole task, can generalize to a variety of real-world insertion tasks involving narrow and negative clearances, all without requiring any fine-tuning.

Abstract

Contact-rich manipulation tasks often exhibit a large sim-to-real gap. For instance, industrial assembly tasks frequently involve tight insertions where the clearance is less than 0.1 mm and can even be negative when dealing with a deformable receptacle. This narrow clearance leads to complex contact dynamics that are difficult to model accurately in simulation, making it challenging to transfer simulation-learned policies to real-world robots. In this paper, we propose a novel framework for robustly learning manipulation skills for real-world tasks using simulated data only. Our framework consists of two main components: the "Force Planner" and the "Gain Tuner". The Force Planner plans both the robot motion and desired contact force, while the Gain Tuner dynamically adjusts the compliance control gains to track the desired contact force during task execution. The key insight is that by dynamically adjusting the robot's compliance control gains during task execution, we can modulate contact force in the new environment, thereby generating trajectories similar to those trained in simulation and narrowing the sim-to-real gap. Experimental results show that our method, trained in simulation on a generic square peg-and-hole task, can generalize to a variety of real-world insertion tasks involving narrow and negative clearances, all without requiring any fine-tuning. Videos are available at https://dynamic-compliance.github.io.

Figures (5)

• Figure 1: Overview of the proposed method. a) In simulation, we use an RL agent to generate data to offline train both the Force Planner and the Gain Tuner. b) In real-world deployment, the Force Planner plans the desired force $f^d$ and robot motion $\Delta x$ to achieve the target return $R$. The Gain Tuner then dynamically adjusts the admittance gains $\mathbf{k}$ to track the desired force.
• Figure 2: Admittance control loop
• Figure 3: Real robot setup and insertion tasks
• Figure 4: a) and c) show the contact force during the rectangular insertion along the $X,Z$ axes, respectively. b) and d) depict the corresponding admittance gains. Our method adaptively adjusts the gains to track the planned contact force generated by the Force Planner. Two baseline methods suffer from the sim-to-real gap and generate noisy gains and contact force.
• Figure 5: The contact force and admittance gains in the $Z$ axis generated by the Gain Tuner while scaling the desired force.