Generalize by Touching: Tactile Ensemble Skill Transfer for Robotic Furniture Assembly

TL;DR

This work addresses the challenge of generalizing long-horizon, contact-rich robotic furniture assembly by introducing TEST, an offline RL framework that fuses tactile feedback with hierarchical skill planning. TEST learns a high-level Skill Transition Model p(z'|z,c) to chain sub-skills and trains intra-skill TEPO policies with hindsight relabeling to overcome sparse rewards, all within a GC-POMDP with multimodal observations. The approach yields about 90% ASR and more than 4x efficiency gains over heuristic policies, and demonstrates robust generalization to unseen furniture configurations as well as resilience to visual disturbances. By integrating tactile sensing and ensemble decision-making into a transformer-based hierarchical pipeline, TEST offers a scalable pathway for transferring contact-rich manipulation skills across varied assembly tasks.

Abstract

Furniture assembly remains an unsolved problem in robotic manipulation due to its long task horizon and nongeneralizable operations plan. This paper presents the Tactile Ensemble Skill Transfer (TEST) framework, a pioneering offline reinforcement learning (RL) approach that incorporates tactile feedback in the control loop. TEST's core design is to learn a skill transition model for high-level planning, along with a set of adaptive intra-skill goal-reaching policies. Such design aims to solve the robotic furniture assembly problem in a more generalizable way, facilitating seamless chaining of skills for this long-horizon task. We first sample demonstration from a set of heuristic policies and trajectories consisting of a set of randomized sub-skill segments, enabling the acquisition of rich robot trajectories that capture skill stages, robot states, visual indicators, and crucially, tactile signals. Leveraging these trajectories, our offline RL method discerns skill termination conditions and coordinates skill transitions. Our evaluations highlight the proficiency of TEST on the in-distribution furniture assemblies, its adaptability to unseen furniture configurations, and its robustness against visual disturbances. Ablation studies further accentuate the pivotal role of two algorithmic components: the skill transition model and tactile ensemble policies. Results indicate that TEST can achieve a success rate of 90\% and is over 4 times more efficient than the heuristic policy in both in-distribution and generalization settings, suggesting a scalable skill transfer approach for contact-rich manipulation.

Figures (6)

• Figure 1: Furniture assembly robot. A natural pipeline of such assembly tasks requires pick, reach, insert, adjust, and thread as the candidate skills to be learned.
• Figure 2: Motivation and challenges in contact-rich robotic assembly problem.
• Figure 3: Diagram of TEST. We tokenize the input sequence with reward-to-go, proprioceptive observation, visual observation, followed by a tactile observations. The input bundle will predict the target pose as the action for the current timestep. At each step, the inputs are aggregated to predict the state at the current timestep.
• Figure 4: Hardware system setup for the experiments.
• Figure 5: Visualization of the embedding of TEST's learned skills t-SNE. We see a clear cluster in each sub-skill, and rollout trajectories in the stool and square table align well.