Dexterous robotic manipulation using deep reinforcement learning and knowledge transfer for complex sparse reward‐based tasks
Problem
Getting a TriFinger robot to dexterously manipulate a cube along a specified positional trajectory under sparse rewards, where traditional control methods struggled.
Why it matters
Dexterous manipulation is a key bottleneck for useful robots in industry and everyday settings.
What the AI did
A DDPG deep reinforcement learning agent using Hindsight Experience Replay and a combined sparse/distance reward learned manipulation strategies in simulation, which were deployed on the real robot.
Results
Won Phase 1 of the Real Robot Challenge 2021 and outperformed all other competition submissions in the final evaluation on the real robot.
Impact
Independent competition evaluation on real hardware confirmed the method beat all other submissions.
Record
| Venue | Expert Systems |
|---|---|
| Published | 2022 |
| Work type | article |
| Authors | Qiang Wang, Francisco Roldán Sánchez, Robert McCarthy, David Córdova Bulens, Kevin McGuinness, Noel Edward O'Connor, Manuel Wüthrich, Felix Widmaier et al. |
| Citations | 17 (OpenAlex, at import time) |
| References | 29 |
| Model or method | Deep Deterministic Policy Gradient (DDPG) with Hindsight Experience Replay (HER) |
| Organization | not stated in source |
| Primary source | https://doi.org/10.1111/exsy.13205 |
| Open access | hybrid |
| Keywords | deep reinforcement learning, DDPG, Hindsight Experience Replay, robotic manipulation, Real Robot Challenge, sim-to-real |
Importance rating
incremental / niche
Dexterous manipulation is a robotics bottleneck, but this is a single competition task (TriFinger cube trajectory) with no stated broader application or scale; strong independent evaluation though.
More in Robotics & Autonomy
| # | Paper | Year | Citations | Importance |
|---|---|---|---|---|
| 1 | Autonomous robotic laparoscopic surgery for intestinal anastomosis | 2022 | 384 |
7
|
| 2 | Solving Rubik's Cube with a Robot Hand | 2019 | 569 |
6
|
| 3 | PaLM-E: An Embodied Multimodal Language Model | 2023 | 346 |
6
|
| 4 | Assistive Navigation Using Deep Reinforcement Learning Guiding Robot With UWB/Voice Beacons and Semantic Feedbacks for Blind and Visually Impaired People | 2021 | 48 |
6
|
| 5 | Towards Deep Radar Perception for Autonomous Driving: Datasets, Methods, and Challenges | 2022 | 154 |
6
|