AnswersPublished13 sources
A two-stage robot learning system pairs vision with touch for precision insertion
At a 0.05 mm clearance, the tactile augmented system achieved 67% insertion success, compared with 52% for imitation plus reinforcement learning without touch. The system used imitation learning for grasping and reaching, then tactile informed reinforcement learning for insertion and recovery from contact errors.
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- At a 0.05 mm clearance, the tactile augmented system achieved 67% insertion success, compared with 52% for imitation plus reinforcement learning without touch.
- The system used imitation learning for grasping and reaching, then tactile informed reinforcement learning for insertion and recovery from contact errors.
- A tactile baseline and torque balanced training helped the robot distinguish insertion contact from grasp loads and learn from difficult contact situations.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: How did Long Cheng’s team at the Institute of Automation of the Chinese Academy of Sciences (CASIA), including Xinpan Meng, Houcheng Li and. Article summary: They separated coarse, vision-guided manipulation from contact-rich precision insertion: imitation learning brought the peg to the hole, while tactile-augmented reinforcement learning learned insertion and recovery from . Topic tags: general, academic, general web, government. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts wi
A small pose error can make a tight-tolerance insertion jam or press too hard against the part. In From Reach to Insert: Tactile-Augmented Precision Assembly under Sub-Millimeter Tolerances, a team from the Chinese Academy of Sciences’ Institute of Automation combined vision-guided reaching with tactile-informed reinforcement learning for the contact-heavy insertion stage. At the tightest tested clearance, 0.05 mm, the approach reached 67% success—better than the no-touch learning comparison, but far from perfect. 1
9
Split the task into reaching and inserting
The team used two learning stages rather than asking one policy to handle the entire task. Diffusion Policy imitation learning controlled grasping randomly placed pegs and moving them toward the hole. Soft Actor-Critic reinforcement learning then handled fine insertion and recovery when contact did not go as planned. 1
This division matters because reaching the opening and correcting an insertion are different problems. Once the peg touches the hole, small alignment errors can change the contact forces and block progress. A learned recovery policy can respond to those interactions instead of simply repeating a demonstrated motion. 1
Use touch to evaluate contact, not just to guide the robot
The method used an asymmetric actor–critic design: the actor was mainly vision-based, while the critic also used tactile information to evaluate actions during learning. In other words, touch helped the system assess contact-rich situations and improve the policy; this is not the same as saying the deployed actor directly relied on tactile input for every action. 1
The researchers also accounted for the forces already present from gripping the peg. They recorded a tactile baseline after grasping and before the peg touched the hole, then compared later readings against that baseline. This helped separate insertion-related force and torque from the load caused by the grasp. 1
For training, they grouped samples by torque and sampled across those groups in a balanced way. This gave the learning process more exposure to important contact and failure examples that might otherwise be less represented. 1
Test the method at increasingly tight clearances
The reported setup used a seven-degree-of-freedom Franka Emika Panda robot arm, a fingertip PaXini PX-6AX GEN3 six-axis tactile sensor, and wrist-mounted and global-view Intel RealSense cameras. Tests covered square, round, hexagonal, L-shaped and triangular holes, with clearances of 1 mm, 0.25 mm and 0.05 mm. 1
| Clearance | Pure imitation success |
|---|---|
| 1 mm | 76% |
| 0.25 mm | 24% |
| 0.05 mm | 0% |
At 0.05 mm, adding reinforcement learning without touch raised success to 52%; the tactile-critic approach with grouped sampling reached 67%, a 15-percentage-point increase. The work also reports reductions of 60% in peak interaction force and 44% in peak torque. 1
9
10
What the results show—and what they do not
The results suggest that touch can make learned insertion more effective and gentler when visual information alone is insufficient to resolve contact. The contribution is not simply adding a tactile signal to a vision policy: it combines staged learning, tactile-informed action evaluation, baseline subtraction and balanced sampling of contact experience. 1
Still, a 67% success rate means the system failed in roughly one-third of the reported attempts at the most demanding clearance. The findings demonstrate improved performance in the tested setup, not reliable insertion in every tight-tolerance task. 1
9
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "A two-stage robot learning system pairs vision with touch for precision insertion"?
At a 0.05 mm clearance, the tactile augmented system achieved 67% insertion success, compared with 52% for imitation plus reinforcement learning without touch.
What are the key points to validate first?
At a 0.05 mm clearance, the tactile augmented system achieved 67% insertion success, compared with 52% for imitation plus reinforcement learning without touch. The system used imitation learning for grasping and reaching, then tactile informed reinforcement learning for insertion and recovery from contact errors.
What should I do next in practice?
A tactile baseline and torque balanced training helped the robot distinguish insertion contact from grasp loads and learn from difficult contact situations.
Sources
arxiv.orgFrom Reach to Insert: Tactile-Augmented Precision Assembly under ...
lefilrobotique.frDossier ICRA / IROS / CoRL - Le Fil Robotique
guandian.hk中科院自動化所程龍團隊提出觸覺增強框架 0.05毫米間隙成功率67% - 觀點網