Research map / Reinforcement learning / Visual and representation RL
Deep Reinforcement Learning for Robotics: A Survey of Real-World Successes
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
A survey of deep RL for robotics that organizes real-world successes by robot competency (locomotion, navigation, manipulation, mobile manipulation, HRI, multi-robot) and rates each paper on a six-level real-world success scale.
Problem
Assessing how mature DRL solutions are for real-world robotics, since most RL milestones have been in simulation or games and physical-world interaction is costly and unsafe.
Approach
A taxonomy along four axes (robot competencies, problem formulation, solution approach, level of real-world success) with a six-level real-world success scale inspired by autonomous driving levels and TRL, applied to filter for papers with at least some real-world validation.
Key results
- Quadruped locomotion is identified as the most mature DRL robotics domain, with policies zero-shot transferred from simulation and deployed on production-level systems at companies such as ANYbotics, Swiss-Mile, and Boston Dynamics.
- Driving-related real-world RL is limited: Jang et al. deployed a cruise control policy on 100 vehicles in a field test, and safety-critical domains such as urban driving and outdoor drone flight have seen few real-world successes.
- Manipulation successes concentrate in constrained tasks (grasping, in-hand manipulation, assembly) that allow a priori reward design and zero-shot sim-to-real transfer, while open-world pick-and-place remains unsolved.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DRL instead of classical feedback controllers such as PID | Quadrotor hovering control · Hovering control | robustness | Helped |
| Use DRL instead of optimization-based controllers | Quadrotor waypoint tracking · Waypoint tracking | tracking error | Hurt |
| Use low-fidelity kinematic-only simulation rather than high-fidelity simulation | Visual legged navigation policies · Indoor and outdoor legged navigation | policy generalizability | Helped |
| Use action-space decomposition for the policy-gradient update | Hierarchical RL for soccer kicking · Soccer kicking | training efficiency | Helped |
| Augment simulation with data-driven residual models of perception and dynamics | Drone-racing RL policy · Drone racing | real-world transfer | Helped |
Surprising findings
- Most commercially deployed navigation systems still rely on classical stacks rather than RL, owing to lack of safety, interpretability, and generalization of RL methods.
- Real-world RL for mobile manipulation is rare because large workspaces make safe exploration hard, so most MoMa work trains in simulation.
Limitations
- Level-of-success assessment is subjective and based only on the authors' reported experiments, which often include a single real-world trial.
- Survey focuses on papers with at least some real-world success, excluding simulation-only work by design.
- Supplementary tables and definitions were partially referenced rather than fully included in the text.
Experimental setup
Metrics: level of real-world success (0-5 scale)
Builds on: Kober et al. robotics RL survey Sunderhauf et al. deep learning for robotics survey Dulac-Arnold et al. real-world RL challenges TRL for machine learning levels of autonomous driving
Closest papers by meaning
- Plan-Seq-Learn: Language Model Guided RL for Solving Long Horizon Robotics Tasks (ICLR 2024 · similarity 0.89)
- Dual RL: Unification and New Methods for Reinforcement and Imitation Learning (ICLR 2024 · similarity 0.89)
- Robot-R1: Reinforcement Learning for Enhanced Embodied Reasoning in Robotics (NeurIPS 2025 · similarity 0.89)
- MENTOR: Mixture-of-Experts Network with Task-Oriented Perturbation for Visual Reinforcement Learning (ICML 2025 · similarity 0.89)
- Stabilizing Contrastive RL: Techniques for Robotic Goal Reaching from Offline Data (ICLR 2024 · similarity 0.88)
- Diffusion Policy Policy Optimization (ICLR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.