From Harmonic Fields to Hierarchical RL: The Innovation Timeline

AMR navigation has evolved across three distinct phases captured in this 80+ patent dataset spanning 1995 to mid-2026. The earliest foundational work — a 1995 German Aerospace Center (Deutsche Forschungsanstalt fur Luft- und Raumfahrt) patent establishing harmonic potential field approaches to collision-free navigation — set the geometric baseline upon which the entire field subsequently built. A US counterpart followed in 1998, and Samsung Electronics filed a SLAM-based walking robot patent in 2012, with Fujitsu contributing landmark detection methods using Mahalanobis distance in 2012–2014.

The mid-stage development period (2016–2020) introduced architectures that persist today. Google LLC’s 2020 WO filing introduced the high-level/low-level policy split — a hierarchical model where a high-level policy selects navigation sub-goals and a separate low-level policy generates precise motor commands — that has become a recurring architecture across the field. iRobot filed coverage robot navigation using beacon-based bounded-area traversal in 2019 (EP), Locus Robotics filed cost-map-based multi-robot navigation planning in 2019 (JP), and SoftBank Robotics Europe disclosed probabilistic obstacle mapping in 2018 (ES).

The dominant share of the dataset — roughly 65 or more filings — falls in the 2021–2026 window. This recent active development period is characterised by the rise of learning-based navigation (deep reinforcement learning, graph neural networks), semantic mapping, multi-robot fleet coordination with shared resource management, and mapless navigation using alternative sensing modalities such as Ultra-Wideband (UWB) radio. According to WIPO, robotics has been among the fastest-growing technology areas in global patent filings over the past decade, a trend clearly reflected in this dataset’s clustering of innovation activity from 2020 onward.

Four Technology Clusters Shaping AMR Navigation in 2026

AMR navigation technology in this dataset spans four interlocking sub-domains: simultaneous localization and mapping (SLAM) and autonomous map-building; path and motion planning; machine learning-driven navigation policies; and multi-robot coordination and fleet management. Each cluster has a distinct maturity profile and IP concentration pattern.

Cluster 1: Geometric Planning with LiDAR Scan Matching and Cost Maps

This is the most mature technical cluster in the dataset. Robots maintain occupancy grid maps or cost maps, match incoming sensor sweeps to known map regions to localise, and plan paths that minimise a cost function balancing distance, obstacle proximity, and path smoothness. Locus Robotics holds multiple filings here: its 2022 multiresolution scan matching patent (ES) builds a pyramid of coincidence maps and exclusion maps at multiple resolutions for computationally efficient LiDAR-based localisation, while its 2021 JP filing generates navigation maps where pixel cost values encode planned robot paths, dynamically reducing collision costs adjacent to each robot’s anticipated trajectory to enable dense multi-robot warehouse operation. Beijing Youzhuju Network Technology (2024, US) integrates global path planning against a SLAM-built map with a local costmap containing real-time obstacle information to generate fine-grained navigation control instructions.

Cluster 2: Machine Learning and Reinforcement Learning Navigation Policies

A fast-growing cluster where navigation behaviours are learned rather than hand-coded. Google LLC’s 2021 US filing trains a high-level policy to select navigation sub-goals and a separate low-level policy to generate precise motor commands achieving those sub-goals while avoiding obstacles, decoupling strategic and tactical navigation. Its 2024 EP filing goes further — automatically learning both a reward function and neural network architecture for a navigation policy network, removing the need for hand-designed reward shaping. Tata Consultancy Services’ 2024 US filing combines a Dynamic Window Approach planner with a novel Next Best Q-learning (NBQ) algorithm whose Q-tree dimension is dynamic and requires no prior dimensionality specification, enabling simultaneous planning and learning. State Grid Jiangsu Electric Power’s 2024 CN filing uses a two-tier deep hierarchical reinforcement learning model trained using LiDAR inputs without requiring a pre-built map.

“Google LLC’s hierarchical policy and deep RL patents (US and EP, active) establish a foundational IP position in end-to-end learned navigation — any commercial product relying on learned high-level/low-level policy separation faces potential exposure.”

Cluster 3: Semantic and Graph-Based Navigation

An emerging cluster that moves beyond purely metric representations to encode semantic knowledge about the environment — object categories, spatial relationships, area identities — enabling goal-directed navigation toward named objects or areas without complete prior maps. Tata Consultancy Services’ 2023 US filing uses a pretrained Graph Neural Network (GNN) over a spatial relationship graph of the indoor environment; the robot computes embedding similarity scores for visible regions to identify the optimal direction toward an out-of-view target object. The same assignee’s 2024 US filing introduces the “AreaGoal” paradigm where the robot selects among room exits based on statistical proximity to a target area and backtracks through a branching spatial graph when the goal is not reached. Baidu USA LLC’s 2023 KR filing uses a pretrained sequence prediction model over a navigation graph to translate natural language commands into sequential single-step robot behaviours, enabling instruction-following navigation without explicit coordinates.

Cluster 4: Autonomous Map-Building (Exploration and SLAM)

Focused on how robots construct maps incrementally in previously unknown environments. Omron Corporation’s 2024 WO filing frames autonomous exploration as a multi-objective optimisation across speed, coverage, and consistency, incorporating loop closure prediction, kinodynamic prioritisation, and a time-based blacklisting table for explored frontier goals. Qualcomm’s 2024 US filing selects among frontier exploration goals by computing a combined cost from trajectory cost and co-visibility cost, minimising total frontier cost to maximise information gain per traversal unit. Tensrai Technologies’ 2025 WO filing integrates SLAM and frontier-based exploration with real-time human override commands, virtual no-go zone definitions, and remote velocity control via a cloud-connected monitoring module — reflecting the broader trend toward human-in-the-loop AMR systems that non-expert operators can guide directly, a direction also noted in standards work by ISO technical committees on service robots.

Who Owns the IP: Assignee Concentration and Geographic Spread

Korea (KR) is the single largest jurisdiction by filing count in this dataset, reflecting strong activity from domestic firms — Navipra, Rainbow Robotics, Naver — and international assignees filing Korean national-phase applications including X Development LLC and Locus Robotics. Europe (EP) is the second most active jurisdiction, dominated by Tata Consultancy Services filings across a broad IP family, followed by the United States (US), Japan (JP), India (IN), and China (CN).

Tata Consultancy Services is the single most prolific filer in this dataset by a substantial margin. Its filings cover semantic navigation, AreaGoal navigation, DWA/NBQ learning hybrids, tele-robot path sensing, and autonomous exploration — almost entirely in EP, US, and IN jurisdictions. This signals a structured, multi-front IP strategy by a major technology services firm, consistent with the kind of systematic IP portfolio building that EPO has identified as characteristic of software-intensive assignees seeking broad jurisdictional coverage.

Google LLC holds a concentrated but high-value cluster around hierarchical and deep reinforcement learning policies filed in WO, US, and EP — suggesting foundational patent positioning rather than broad volume. Locus Robotics holds approximately five filings explicitly targeting warehouse environments. Seegrid Corporation contributes approximately four filings focused on fleet-level coordination via shared resource (space) management. X Development LLC holds approximately five filings across KR. Qualcomm holds approximately three filings in US, IN, and WO jurisdictions focused on frontier-based map building.

Innovation is moderately concentrated: the top five assignees account for the majority of filings in this dataset, though a long tail of single-patent entities — Tensrai Technologies, TMRW Foundation, Wipro, Proxima Robotics, Harbin Institute of Technology — indicates distributed academic and startup contributions that are characteristic of an emerging field, a pattern also documented in OECD research on frontier technology patent ecosystems.

Six Emerging Directions Defining the Next Wave

The most recent filings in this dataset (2024–2026) reveal six converging frontier directions that are reshaping the AMR navigation IP landscape and creating near-term competitive differentiation opportunities.

1. Mapless Navigation via UWB: Two 2025 KR filings from Dongeui University Industry-Academic Cooperation Foundation propose anchor-based UWB infrastructure as a localisation substrate, removing dependence on LiDAR maps entirely and making the approach suitable for GPS-denied industrial halls. Filings are sparse and jurisdiction coverage is thin — representing a near-term opportunity for first-filer advantage, particularly in CN and US jurisdictions.

2. Digital Twin / Virtual World-Guided Navigation: TMRW Foundation IP SARL’s 2023 EP filing proposes computing navigation routes against a 3D virtual replica of the real environment with geolocation-synchronised virtual objects, effectively using a digital twin as the planning substrate. This approach is early-stage and jurisdiction-limited (EP, IN), and currently has limited blocking IP against it in the US.

3. Hierarchical Attention and Foundation Model Integration: Soochow University’s 2026 CN filing fuses spatial semantic image features, dynamic pedestrian trajectory features, and positional goal features via hierarchical attention, training an actor-critic policy via deep reinforcement learning. Naver Corporation’s 2025 KR filing uses a backbone neural network trained across multiple navigation tasks with adapter layers for generalisation to new task configurations — a direction consistent with the broader foundation model paradigm documented by IEEE in recent robotics and automation proceedings.

4. Connectivity-Aware Path Selection for Tele-Robots: Tata Consultancy Services’ 2026 EP filing introduces a Log-Normal Shadowing Model (LNSM) for predicting radio signal strength along candidate future paths, enabling tele-robots to select paths that preserve communication connectivity — a critical constraint for remotely operated robots in signal-degraded environments.

5. Perception-Uncertainty-Driven Route Switching: Aurora Flight Sciences (a Boeing subsidiary) formalises the transition from a pre-planned global route to a dynamically generated local route based on measured perceptual uncertainty in its 2025 EP filing, providing a principled framework for handling sensor degradation during missions.

6. Natural Language and Human-in-the-Loop Interfaces: Multiple recent filings embed natural language command parsing (Baidu USA LLC, 2023, KR) and human feedback integration into autonomous mapping (Tensrai Technologies, 2025, WO), reflecting the move toward instruction-following AMRs that non-expert users can task directly.

Strategic Implications for R&D and IP Teams

Five strategic conclusions emerge from this patent landscape analysis that are directly actionable for R&D directors, IP counsel, and product teams working on AMR navigation systems.

TCS holds a disproportionate share of semantic and area-goal navigation IP across EP, US, and IN jurisdictions. R&D teams developing indoor service robots should assess freedom-to-operate around the AreaGoal, GNN-semantic, and DWA/NBQ learning families before committing to competing architectures. The breadth of TCS’s multi-front coverage — spanning sensing, planning, learning, and tele-operation — is unusual for a technology services firm and suggests a deliberate IP monetisation strategy.

Google LLC’s hierarchical policy and deep RL patents (US and EP, active) establish a foundational IP position in end-to-end learned navigation. Any commercial product relying on learned high-level/low-level policy separation faces potential exposure; licensing or design-around is an early strategic decision that should be made before significant engineering investment is committed.

Mapless navigation (UWB, deep RL without prior maps) is an underdeveloped but rapidly growing whitespace. Filings are sparse, jurisdiction coverage is thin, and the enabling technologies — UWB infrastructure, hierarchical RL — are maturing. This represents a near-term opportunity for first-filer advantage, particularly in CN and US jurisdictions where current coverage is minimal.

Fleet coordination and shared resource management is emerging as a distinct IP sub-domain. As AMR deployments scale from single-unit to multi-unit fleets in warehouses and manufacturing, collision arbitration, congestion-aware routing, and resource (space) locking are becoming commercially critical and patent-contested — with Seegrid, Navipra, and Naver all active in this space.

The digital twin / virtual world navigation architecture is early-stage and jurisdiction-limited. TMRW Foundation’s EP and IN filings currently face limited blocking IP in the US. If validated operationally, this approach could become foundational infrastructure for next-generation AMR deployment in complex, frequently changing environments — and early movers in the US and CN jurisdictions would face minimal prior art obstacles today.

“Fleet coordination and shared resource management is emerging as a distinct IP sub-domain — as AMR deployments scale from single-unit to multi-unit fleets, collision arbitration and congestion-aware routing are becoming commercially critical and patent-contested.”