Zekai Jin1, Huiguang Wang1, Yihong Tang1, Zhen Dong2,3, Chen Feng4, and Yi Shao1,5,*
1Department of Civil Engineering, McGill University, Montreal, QC, Canada
2Department of Computer Science, University of California, Santa Barbara, Santa Barbara, CA, USA
3NVIDIA Corporation, Santa Clara, CA, USA
4Department of Mechanical and Aerospace Engineering, Tandon School of Engineering, New York University, NY, USA
5Department of Civil Engineering, University of British Columbia, Vancouver, Canada
*Corresponding author: yi.shao@ubc.ca
DRL4CONBOTS Evidence Atlas
Preview interactive atlas | Browse static page source | Copy citation metadata
A curated companion repository for Deep Reinforcement Learning for Construction Robotics (DRL4CONBOTS).
This repository accompanies the manuscript under review in Automation in Construction, "Deep Reinforcement Learning for Construction Robotics: A System-Level Taxonomy and Evidence Map toward Real-World Readiness".
DRL-enabled construction robotics is moving from isolated task demonstrations toward system-level questions about authority, runtime assurance, validation exposure, and deployment-relevant evidence. This repository tracks representative papers, taxonomy figures, and evidence patterns for researchers, practitioners, and students working at the intersection of construction automation, robot learning, and field robotics.
Contact: zekai.jin@mail.mcgill.ca
| Item | Current release |
|---|---|
| Manuscript status | Under review in Automation in Construction |
| Evidence base | 152 coded construction-robotics instances |
| Primary DRL synthesis | 75 primary DRL instances across five competency regimes |
| Framework | A1-A5 competency, learning formulation, runtime authority, runtime assurance, and validation exposure |
| Core diagnostic | 35/75 primary instances are AM0 + EVL L0, while 0 reach sustained workflow-integrated deployment |
The repository includes a GitHub Pages-ready interactive atlas in docs/index.html. The atlas provides a high-level entry point for readers who want to scan the framework, inspect readiness diagnostics, filter representative papers, and browse core figures before reading the manuscript.
Graphical overview of the DRL4CONBOTS evidence framework for deployment-relevant DRL in construction robotics.
- At a glance
- Interactive Evidence Atlas
- News and Updates
- Corpus and Screening
- Framework and Taxonomy
- Evidence Diagnostic
- Representative Papers by Category
- Evidence Trends
- Primary Evidence Regimes
- Citation
- License & contact
- [June 2026] Repository launch:
DRL4CONBOTSis prepared as the public companion repository for the manuscript under review in Automation in Construction. - [June 2026] Evidence map released: The current repository tracks 152 coded construction-robotics instances, including 75 primary DRL instances.
- [June 2026] Core finding: Across the 75 primary DRL instances, 35 are simultaneously AM0 and EVL L0, while 0 reach sustained workflow-integrated deployment.
The review uses a report-level PRISMA screening process before coding system-instance evidence for taxonomy and readiness analysis.
DRL4CONBOTS is a system-level evidence map for deep reinforcement learning in construction robotics. Instead of ranking algorithms by task success alone, it interprets what each reported robotic system can credibly claim based on its task regime, learning formulation, runtime authority, runtime assurance, and validation exposure.
The review uses five coupled dimensions:
- A1. Competency regime: earthwork, assembly, lifting, additive/surface processing, or navigation/logistics.
- A2. Learning formulation: observation interface, action abstraction, decision formalism, objective specification, and training pathway.
- A3. Runtime authority: who senses, proposes, arbitrates, and executes robot behavior.
- A4. Runtime assurance: whether execution-time safeguards, constraints, monitors, or fallback layers are disclosed.
- A5. Validation exposure: where behavior is tested, from simulation-only evaluation to sustained field deployment.
Five construction-robotics competency regimes used to organize task-specific evidence and deployment bottlenecks.
Methodological translation logic: reported task outcomes are interpreted as bounded deployment-relevant claim scope through the five-axis framework.
A cross-axis view of runtime-assurance disclosure and validation exposure across competency regimes.
Runtime assurance (A4) against validation exposure (A5), with validation-exposure composition by competency regime.
Selected papers are grouped by the five competency regimes used in the DRL4CONBOTS framework.
Earthwork and Material Processing - 36 primary instances
Function: Learn excavation, grading, loading, material handling, and machine-level interaction where soil, rock, traction, tool load, and material-state evolution dominate claim scope.
- Near-operational material handling
- Large-Scale Robotic Material Handling: Learning, Planning, and Control. Spinelli, F. A. et al. [IEEE TFR 2026].
- Highlight: The only primary instance reaching EVL L4 in the current synthesis; combines learning, planning, real-machine material handling, and disclosed runtime constraints.
- Large-Scale Robotic Material Handling: Learning, Planning, and Control. Spinelli, F. A. et al. [IEEE TFR 2026].
- Full-scale excavation and loading
- Reinforcement Learning-Based Bucket Filling for Autonomous Excavation. Egli, P. et al. [IEEE TFR 2024].
- Highlight: Connects randomized simulation with full-size excavation and bucket-filling evidence.
- Automatic Loading of Unknown Material with a Wheel Loader Using Reinforcement Learning. Eriksson, D. et al. [ICRA 2024].
- Highlight: Studies online material adaptation on a 24-tonne wheel loader under real-machine loading conditions.
- Reinforcement Learning-Based Bucket Filling for Autonomous Excavation. Egli, P. et al. [IEEE TFR 2024].
- Safety-aware excavation control
- Safe reinforcement learning for tracking control of uncertain hydraulic excavators. Chen, K. et al. [Nonlinear Dynamics 2025].
- Highlight: Uses safe-RL tracking logic for uncertain hydraulic excavator control and is one of the few AM3-coded primary instances.
- Safe reinforcement learning for tracking control of uncertain hydraulic excavators. Chen, K. et al. [Nonlinear Dynamics 2025].
Structural Assembly and Installation - 18 primary instances
Function: Learn tolerance-sensitive placement, insertion, joining, tactile correction, and installation where small pose errors can become jamming, wedging, surface damage, or contact-mode failures.
- Contact-rich timber and architectural assembly
- Robotic assembly of timber joints using reinforcement learning. Apolinarska, A. A. et al. [Automation in Construction 2021].
- Highlight: Combines simulation training, force/torque sensing, and physical lap-joint validation.
- Robotic architectural assembly with tactile skills: Simulation and optimization. Belousov, B. et al. [Automation in Construction 2022].
- Highlight: Shows why tactile simulation fidelity matters for modular assembly and contact-rich construction tasks.
- Robotic assembly of timber joints using reinforcement learning. Apolinarska, A. A. et al. [Automation in Construction 2021].
- Installation and tactile transfer
- Visual-tactile learning of robotic cable-in-duct installation skills. Duan, B. et al. [Automation in Construction 2025].
- Highlight: Connects SAC control, tactile sim-to-real alignment, and physical cable-in-duct trials.
- Training of construction robots using imitation learning and environmental rewards. Duan, K. et al. [Computer-Aided Civil and Infrastructure Engineering 2025].
- Highlight: Combines imitation learning and environmental rewards for installation-oriented construction robot training.
- Visual-tactile learning of robotic cable-in-duct installation skills. Duan, B. et al. [Automation in Construction 2025].
Material Placement and Lifting - 8 primary instances
Function: Learn crane, hoist, lift-planning, and suspended-payload behavior where delayed oscillatory dynamics, swept-volume risk, payload variability, and site governance shape readiness claims.
- Construction-representative crane control
- Autonomous construction framework for crane control with enhanced soft actor-critic algorithm and real-time progress monitoring. Xiao, Y. et al. [Computer-Aided Civil and Infrastructure Engineering 2025].
- Highlight: The strongest disclosed lifting-regime validation anchor in the primary set.
- Autonomous construction framework for crane control with enhanced soft actor-critic algorithm and real-time progress monitoring. Xiao, Y. et al. [Computer-Aided Civil and Infrastructure Engineering 2025].
- Crane dynamics and stabilization
- Controlling a double-pendulum crane by combining reinforcement learning and conventional control. Eaglin, G. et al. [ACC 2023].
- Highlight: Demonstrates why hybrid conventional-control-plus-RL structures can outperform RL-only control in crane dynamics.
- Online reinforcement learning with passivity-based stabilizing term for real time overhead crane control without knowledge of the system model. Zhang, M. et al. [MSSP 2022].
- Highlight: Uses passivity-based stabilization to structure online RL for overhead crane control.
- Controlling a double-pendulum crane by combining reinforcement learning and conventional control. Eaglin, G. et al. [ACC 2023].
- Lift planning
- Reinforcement learning-based simulation and automation for tower crane 3D lift planning. Cho, S. et al. [Automation in Construction 2022].
- Highlight: Evaluates tower-crane lift planning in a real-scale virtual site and compares simulated planning behavior with field-related traces.
- Reinforcement learning-based simulation and automation for tower crane 3D lift planning. Cho, S. et al. [Automation in Construction 2022].
Additive Manufacturing and Surface Processing - 2 primary instances
Function: Learn process control and path planning where material rheology, tool wear, cumulative geometry deviation, and irreversible process defects shape the validation burden.
- Robotic additive manufacturing
- Autonomous robotic additive manufacturing through distributed model-free deep reinforcement learning in computational design environments. Felbrich, B. et al. [Construction Robotics 2022].
- Highlight: Uses distributed model-free DRL for robotic additive manufacturing in computational design environments.
- Autonomous robotic additive manufacturing through distributed model-free deep reinforcement learning in computational design environments. Felbrich, B. et al. [Construction Robotics 2022].
- Concrete 3D printing
- Reinforcement learning-based continuous path planning and automated concrete 3D printing of complex hollow components. Wang, X. et al. [Automation in Construction 2025].
- Highlight: Optimizes continuous fill-path sequencing for concrete 3D printing of complex hollow components.
- Reinforcement learning-based continuous path planning and automated concrete 3D printing of complex hollow components. Wang, X. et al. [Automation in Construction 2025].
Navigation, Layout, and Logistics Support - 11 primary instances
Function: Learn mobility, routing, worker-aware planning, layout support, and logistics decisions where human co-presence, congestion, occlusion, handover, and governance determine claim scope.
- Safety-constrained construction HRC
- Safety-constrained Deep Reinforcement Learning control for human-robot collaboration in construction. Duan, K. et al. [Automation in Construction 2025].
- Highlight: One of the few navigation/HRC instances with explicit safety-constrained runtime evidence.
- Safety-constrained Deep Reinforcement Learning control for human-robot collaboration in construction. Duan, K. et al. [Automation in Construction 2025].
- Scene-graph and bulldozer navigation
- Deep reinforcement learning coupled with topological scene graph for dynamic path planning of autonomous bulldozer in complex earthwork construction. Gao, H. et al. [Automation in Construction 2026].
- Highlight: Combines topological scene graphs and DRL for dynamic bulldozer path planning in complex earthwork settings.
- Deep reinforcement learning coupled with topological scene graph for dynamic path planning of autonomous bulldozer in complex earthwork construction. Gao, H. et al. [Automation in Construction 2026].
- Worker-aware planning
- Prediction-based path planning for safe and efficient human-robot collaboration in construction via deep reinforcement learning. Cai, J. et al. [Journal of Computing in Civil Engineering 2023].
- Highlight: Combines worker-location prediction with DQN path planning for construction HRC scenarios.
- Prediction-based path planning for safe and efficient human-robot collaboration in construction via deep reinforcement learning. Cai, J. et al. [Journal of Computing in Civil Engineering 2023].
The current evidence base shows that real-world readiness cannot be inferred from algorithm labels, hardware presence, or field-like demonstrations alone.
Key findings from the primary synthesis set:
- Authority-assurance decoupling: 45/75 primary instances are AM0, and 60/75 remain within AM0-AM1.
- Simulation without safeguards: 35/75 primary instances are simultaneously AM0 and EVL L0.
- Exposure-assurance divergence: 20/75 primary instances reach EVL L3 or higher, but only 1 reaches EVL L4 and 0 reach EVL L5.
| Regime | Primary instances | Main bottleneck |
|---|---|---|
| Earthwork and Material Processing | 36/75 | Long-horizon exposure to soil variability, muck, resistance discontinuities, and intervention histories |
| Structural Assembly and Installation | 18/75 | Tolerance variation, contact-rich disturbances, access constraints, and intervention reporting |
| Material Placement and Lifting | 8/75 | Workflow-coupled lift validation, swing handling, disturbance coverage, and governance |
| Additive Manufacturing and Surface Processing | 2/75 | Scale-dependent process physics, material rheology, and end-to-end process assurance |
| Navigation, Layout, and Logistics Support | 11/75 | Mixed-traffic exposure, handover, communication stress, recovery, and site governance |
The manuscript is currently under review in Automation in Construction. Citation metadata will be updated after preprint, acceptance, or journal publication.
@misc{jin2026drl4conbots,
title = {Deep Reinforcement Learning for Construction Robotics: A System-Level Taxonomy and Evidence Map toward Real-World Readiness},
author = {Jin, Zekai and Wang, Huiguang and Tang, Yihong and Dong, Zhen and Feng, Chen and Shao, Yi},
year = {2026},
howpublished = {Companion repository for a manuscript under review in Automation in Construction},
url = {https://github.com/ZekaiJ/DRL4CONBOTS}
}Released under the MIT license. Questions, feedback, or collaboration: zekai.jin@mail.mcgill.ca