An Actor-Critic deep-reinforcement-learning (DRL) model published in IEEE Xplore on 23 June 2026 cut path length to 14.6 m, raised success rate to 97.8%, and converged in 430 iterations, beating A*, DQN, and PPO baselines on the same complex-environment benchmark [S1].
Parallel work in the Robot Learning journal (DOI 10.55092/rl20260005) introduced a Path Planning Transformer (PPT) that imitates an improved RRT* with reduced random map size (IRRT*-RRMS), runs on a standard laptop with MATLAB, ROS, and Gazebo, and re-plans smoother trajectories for two-robot teams after LiDAR-detected obstacles or right-of-way conflicts [S5]. The two papers, plus a 2026 review on autonomous-mobile-robot path planning, frame AI-driven path planning as a layered stack rather than a single algorithm choice.
What "AI-driven path planning" means in 2026
AI-driven path planning uses learned policies or sequence models to generate, evaluate, or re-plan trajectories, instead of (or on top of) classical graph search and sampling planners. The 2026 review defines the field as planners that "change their path dynamically in case of real-time alterations, e.g., aisles being blocked," with AI-oriented planning as the response layer to map updates and other-agent motion [S3].
In the IEEE paper, the AI is an Actor-Critic network that operates inside a Markov decision process, with a multidimensional state space, multimodal action space, and a multi-objective coupled reward that weights path quality against obstacle avoidance; the network jointly optimises policy and value functions during training [S1]. In the Robot Learning paper, the AI is a Transformer that maps occupancy grids directly to smooth paths after being supervised on expert trajectories from IRRT*-RRMS, with a right-of-way rule added to handle multi-robot interactions without centralised coordination [S5]. Both approaches share one principle: classical planners generate the training signal, the neural model generalises the behaviour, and on-board perception (LiDAR, depth cameras, occupancy maps) feeds the runtime state [S1][S5].
Algorithm landscape: classical vs learning-based vs hybrid
Three planner families dominate 2026 deployments. Classical search and sampling (A*, D*, RRT, RRT*) are deterministic, well-understood, and easy to certify; DRL planners (DQN, PPO, Actor-Critic) learn policies that generalise across map distributions; Transformer planners (PPT, decision Transformers, diffusion planners) treat path generation as sequence or token prediction and excel at smoothness and replanning latency [S1][S5].
The IEEE benchmark quantifies the gap: the proposed Actor-Critic model drove path length down to 14.6 m and success rate to 97.8% versus the A*, DQN, and PPO baselines on the same test, with convergence reached at 430 iterations [S1]. The Robot Learning study reports a different, complementary advantage: PPT produced smoother paths with fewer turns than A* and RRT* in two-robot replanning tests, even when the classical planner found a shorter nominal route [S5]. Practical takeaway from the 2026 evidence: learning-based planners do not replace A* and RRT*, they sit on top of them as a smoothness and re-plan layer, which is the same pattern already used for SCARA robot pick-and-place and collaborative robot safety zones.
Decision criteria: when to use which planner
Selection boils down to four engineering criteria: environment structure, replanning frequency, compute budget, and certification burden. Static warehouses with fixed aisles favour A* with a D* lite fallback because the map rarely changes and the algorithm is straightforward to validate; the 2026 review notes that AI-oriented methods target the "real-time alterations" case where classical pipelines struggle [S3].
Dynamic, mixed-fleet floors with frequent re-routing are where the IEEE Actor-Critic and PPT results pay off, with the IEEE model showing 97.8% success and 430-iteration convergence on a complex test, and the PPT model running on a standard laptop using MATLAB, ROS, and Gazebo rather than a GPU server [S1][S5]. Compute-constrained mobile platforms should default to lightweight learning layers (Transformer imitation of RRT*) rather than full PPO training on the robot, since on-line PPO remains expensive; mobile robot vendors already use this split. Safety-critical cells with formal validation requirements (for example, fenced industrial robot work envelopes near humans) should keep a certified classical planner as the primary path generator and use a learning model only as a re-route suggestion under supervisory control.
Industrial use cases: warehouses, factories, and shared spaces
The Robot Learning study targets two-robot teams in 2D occupancy grids, using LiDAR to detect other robots and unexpected obstacles, then injecting a virtual obstacle that enforces a preferred passing direction under a modified right-of-way rule [S5]. This is the same problem class faced by AMR robot fleets in e-commerce fulfilment, where aisle blockages, fallen totes, and human-occupied zones force re-plans every few seconds.
The 2026 review adds factory automation and service robotics as the deployment targets, and notes that AI-oriented planners let robots "change their path dynamically" when aisles or stations become unavailable, a frequent event in mixed-model production [S3]. The 2025 AI-assisted CAM work referenced in adjacent coverage shows the same pattern of AI cutting cycle-time in CNC, reinforcing that path-quality optimisers are now common across both motion and machining domains. For articulated arms, a learning-based planner can sequence joint waypoints that an articulated robot controller then interpolates; for mobile bases, the same model can output velocity and steering commands directly from the occupancy grid.
Known limits: training data, sim-to-real, and scaling
The Robot Learning authors are explicit about scope: the current study focuses on two-robot scenarios in 2D environments, and future work will extend to larger teams and 3D voxel-based maps [S5]. The IEEE paper likewise tests on a single complex environment rather than a benchmark suite, so the 97.8% success rate and 14.6 m path length are not portable claims across factories [S1].
Other practical constraints: DRL training needs thousands of episodes, IRRT*-RRMS supervision in PPT needs thousands of expert trajectories generated by an improved RRT* with reduced random map size, and both pipelines assume accurate occupancy maps from LiDAR or depth fusion [S1][S5]. Classical planners can be more brittle in dynamic settings: the Robot Learning paper notes that "classical planners such as A* or RRT* are reliable but can struggle to re-plan smoothly in dynamic environments, especially when multiple robots are involved," which is exactly the failure mode the Transformer addresses [S5]. Conversely, learning models can produce shorter, more dynamic paths than A* but introduce opaque decisions that complicate safety cases; the right deployment is hybrid, with a classical planner as the certified baseline and a learned re-planner as the speed layer.
Tooling, standards signals, and where to track progress
Tooling is converging on open stacks. The PPT system runs on MATLAB, ROS, and Gazebo with two real mobile robots, and trains the Transformer on thousands of automatically generated IRRT*-RRMS expert paths [S5]. The IEEE DRL model uses an Actor-Critic architecture inside a Markov decision process framework with a multi-objective reward, validating on path length, planning time, and success rate against A*, DQN, and PPO baselines [S1].
There is no single IEC or ISO standard dedicated to AI-driven robot path planning as of 2026, but ISO 13849 (safety of machinery control systems) and ISO 10218 (industrial robot safety) remain the governing frameworks for any planner that can affect a safeguarded space. For shared human-robot zones, vendors typically wrap a learned re-planner inside a certified classical safety controller rather than certifying the network itself; the December 2026 ICARPP conference in Istanbul lists 11 tracks, including "Path Planning Techniques in Dynamic Environments" and "Optimization Techniques in Robotics Path Planning," which is a signal of where peer-reviewed convergence is heading [S4]. Two trackable signals: ICARPP-26 paper acceptance outcomes (deadline 21 November 2026) and any release of PPT or Actor-Critic reference code on ROS or GitHub, both of which will move the field from single-paper benchmarks to multi-site reproducibility.
For related coverage, see Open-path gas detector alignment: peak signal procedure and commissioning checks.