SymbioticFlow: Heterogeneous Multi-Robot Cargo Handling for Railway Hubs

AI Quick Summary

SymbioticFlow introduces a radically new approach to automating railway hub logistics: a heterogeneous multi-robot system where specialized robots collaborate through an agentic AI platform to handle the full cargo flow (from rail wagon to sorting belt) without any infrastructure modifications. A wheeled humanoid robot (for this study: Galaxea R1 Pro) provides AI-driven parcel manipulation at the conveyor interface, while a customized wheeled-legged quadruped (for this study: DEEPRobotics LYNX M20 Pro) autonomously transports trolleys across the unstructured hub terrain. Both robots are coordinated by Binabik AI's Agentic Operating System, which uses vision-language models to reason about tasks in real time. The consortium (Binabik AI + ZHAW) has both robots in hand and the AI software stack operational. This Proof of Concept validates a single robotic cell; the architecture scales to full-hub and multi-hub deployment through straightforward replication. 1. Executive Summary SymbioticFlow introduces a radically new approach to automating railway hub logistics: a heterogeneous multi-robot system where specialized robots collaborate through an agentic AI platform to handle the full cargo flow (from rail wagon to sorting belt) without any infrastructure modifications. A wheeled humanoid robot (Galaxea R1 Pro) provides AI-driven parcel manipulation at the conveyor interface, while a wheeled-legged quadruped (DEEPRobotics LYNX M20 Pro) autonomously transports trolleys across the unstructured hub terrain. Both robots are coordinated by Binabik AI's Agentic Operating System, which uses vision-language models to reason about tasks in real time. The consortium (Binabik AI + ZHAW) has both robots in hand and the AI software stack operational. This Proof of Concept validates a single robotic cell; the architecture scales to full-hub and multi-hub deployment through straightforward replication. After this proof of concept, the main limitation stopping wider deployment and scalability will be hardware. However, humanoid robots and mobile manipulators with 20-30kg payload will enter the market next year, while custom robots can be built to meet these specifications. The approach proposed in this project (Agentic OS and software for multi-robot coordination without infrastructure modifications) applies to any mobile robot and mobile manipulation platform. We simply use a wheeled humanoid and quadruped because they are sufficient for the proof of concept in this project, while the solution remains hardware-agnostic. 2. Background and Problem Statement Every working night, Planzer processes tens of thousands to hundreds of thousands of parcels, unloading rail wagons and feeding heterogeneous cargo onto sorting belts for last-mile distribution. This operation is mostly manual: workers navigate railway tracks and through the trains to pull trolleys from wagons, then unload parcels one by one onto conveyor belts, all under extreme time pressure. The environment is fundamentally hostile to conventional automation. Railway hubs are rented SBB infrastructure where no permanent modifications are usually permitted. Clearances are tight and the environment is cluttered. Layouts differ significantly between hubs (Zurich, Bern, and others each present different spatial configurations). There are no standardized loading patterns: cargo varies in size, weight, shape, and fragility. This is why traditional approaches fail here: Fixed conveyor extensions require infrastructure changes that SBB constraints prohibit, and cannot adapt to varying hub layouts. Standard warehouse AMRs need flat, clean floors and structured environments for localization and robust navigation, while they cannot always traverse rails, gaps, or indoor-outdoor transitions. Industrial robotic arms require structured, predictable environments with standardized cargo, almost the opposite of a railway hub. Meanwhile, Europe faces a shortage of 745,000 logistics workers. The overnight shifts involved are physically demanding, ergonomically hazardous, and increasingly difficult to staff. The need for automation is urgent, but the solution must work within the constraints of these environments, not demand that the environments change. 3. Proposed Solution: The SymbioticFlow Architecture 3.1 Design Principle: The Right Robot for the Right Task Rather than forcing a single robot type to handle everything poorly, SymbioticFlow decomposes the cargo flow into two specialized roles, each matched to the morphology and capabilities of a specific robot platform, coordinated by an intelligent Physical Agentic AI layer. 3.2 The Dexterous Node: Wheeled Humanoid Robot The Galaxea R1 Pro is stationed at the conveyor belt interface, where it performs the cognitively demanding task of identifying, grasping, and placing heterogeneous parcels. Its wheels allow it to reach trolleys in different places, compared to a fixed robotic arm with limited workspace dimensions. Key characteristics include: AI-driven perception: Using Vision-Language Models (VLMs), the robot visually assesses each parcel's type, size, orientation, and potential fragility before grasping, enabling zero-shot adaptation to parcel types never seen during training. Dual-arm manipulation: Two 6-DoF arms with parallel grippers (26 DoF total) execute precise pick-and-place operations. The wheeled base provides micro-positioning adjustments around the trolley. Vision-Language-Action models are trained to handle the packages. Capacity: Rated for 7 kg dual-arm payload (10 kg maximum), covering a large number of parcel weight classes. The 170 cm height and 0–200 cm vertical operating range match the full trolley unloading envelope. The solution is easily deployable to robots with higher payload that will reach the market in the next 6-12 months. Onboard compute: NVIDIA Jetson AGX Orin enables real-time VLM inference with end-to-end latency under 200 ms. 3.3 The All-Terrain Tugger: Wheeled-Legged Quadruped The available LYNX M20 operates as the logistics runner, autonomously transporting trolleys between rail wagons and the sorting station. The robot will be modified with a hitch specific to couple with Planzer's standard trolleys. Terrain traversal: Its hybrid wheeled-legged design switches seamlessly between wheeled mode (fast, efficient on flat surfaces) and legged articulation (stepping over rails, navigating gaps, handling uneven terrain). It can handle light slopes and small gaps. Trolley towing: Equipped with a mechanical hitch mechanism, the LYNX crouches, slides under the trolley base, locks in, and tows it. Rolling resistance on wheeled trolleys is low: the challenge is the attachment/detachment cycle and terrain navigation, not raw pulling force. All-conditions operation: IP66-rated, operational from -20 to 55 degrees C, with bidirectional lighting. If needed, it can be exposed to weather. Perception: Dual 96-line LiDAR units (360x90 degree FOV) and wide-angle cameras provide robust SLAM-based navigation through cluttered, dynamic hub environments. 3.4 The Agentic Operating System: Binabik AI's Software Platform The robots are coordinated by Binabik's three-layer Agentic OS: Cognition Layer: VLM/LLM-based reasoning decomposes high-level goals ("unload wagon 7") into robot-specific task sequences, adapting in real time to unexpected situations (blocked paths, unusual parcels, equipment changes). Execution Layer: A Model Context Protocol (MCP) interface dispatches specialized skills: VLA policies for parcel manipulation, reinforcement learning (RL) policies for hitch engagement, ROS 2/Nav2 modules for navigation. Control Layer: Safety-aware whole-body control ensures compliant, predictable robot behavior in shared human-robot workspaces. This architecture is inherently modular: new skills can be added for additional tasks (sorting, van loading, inspection) without redesigning the system. The same Agentic OS runs across different robot hardware (already validated on Galaxea, Unitree, and DEEPRobotics platforms), making it a natural building block for broader end-to-end logistics automation. 4. POC Scope and Deployment Path This project validates a single robotic cell: one Galaxea R1 Pro unloading parcels from one trolley onto a conveyor belt, with one LYNX M20 delivering and retrieving trolleys. The POC measures cycle time, grasp success rate, and navigation reliability in a real Planzer hub environment. For full-hub deployment, the cell is replicated — 50 to 100 parallel cells coordinated by the fleet management layer of the Agentic OS. Larger quadruped platforms or electric tug attachments can substitute or augment the LYNX for heavier trolleys. Next-generation humanoids (Unitree H2, available at ZHAW) extend the parcel weight range. The POC validates the unit economics and technical feasibility that underpin this scaling path. 5. Objectives O1: Demonstrate autonomous unloading of heterogeneous parcels (up to 7 kg) from a standard Planzer trolley onto a conveyor belt, achieving a cycle time of less than 90 seconds per parcel. O2: Demonstrate autonomous LYNX M20 trolley towing through a representative railway hub section, including rail crossing, narrow passages, and the hitch/unhitch cycle. O3: Validate the Binabik Agentic OS as a real-time coordination layer managing task allocation between the two heterogeneous robots. O4: Characterize the parcel weight and size distribution at a real Planzer hub to quantify the coverage of current robot capabilities and inform next-generation system design. O5: Produce a validated cost model and scaling roadmap for full-hub and multi-hub deployment. 6. Implementation Plan (6 Months) Phase 1: Lab Integration (Month 1–2) Visit Planzer railway hub with robots for initial exploration of capabilities with teleoperation Construct mock trolley and conveyor belt interface at ZHAW/Binabik lab Deploy Galaxea R1 Pro with initial VLA model for parcel grasping (leveraging Binabik's existing pick-and-place skill library) Deploy LYNX M20 for indoor navigation and hitch mechanism testing Integrate Agentic OS orchestration: task dispatch, robot status monitoring, error recovery Deliverable: Both robots operating individually in lab; initial integrated handoff demonstrated Phase 2: Capability Development (Month 2–4) Fine-tune VLA models on Planzer-representative parcel types (varied sizes, shapes, weights, packaging materials) Develop robust VLM-based grasp planning for novel and irregularly shaped parcels Optimize LYNX navigation for rail-crossing and uneven terrain (simulation first via Isaac Sim, then physical mockup) Refine multi-robot coordination: LYNX signals trolley delivery, Galaxea begins unloading sequence, empty trolley triggers retrieval Deliverable: Integrated two-robot demo in lab with measured cycle times per parcel Phase 3: Planzer Test Environment (Month 4–6) Deploy at real Planzer hub (test environment access provided as part of the award) Conduct systematic trials: measure parcel throughput, grasp success rate by parcel type, failure mode analysis Collect real parcel distribution data (weight, dimensions, fragility categories) Demonstrate live operation to Planzer stakeholders Deliverable: Validated POC with quantitative performance data in operational environment Phase 4: Analysis and Dissemination (Month 6) Compile results into scaling analysis and cost model Present findings at Innovate Mobility co-creation workshop (November 2026) Prepare follow-on funding proposals (Innosuisse, Horizon Europe) Submit results to leading robotics venues (ICRA, IROS) Deliverable: Final report, scaling roadmap, academic publication draft, demo video 7. Novelty and Advancement of the State of the Art Current logistics automation operates at two extremes: highly structured warehouse systems that require purpose-built facilities, and experimental humanoid robots that remain confined to controlled laboratory or factory settings. Neither approach works in the unstructured, constrained, and variable environments of railway logistics hubs. SymbioticFlow advances the state of the art in three ways: Heterogeneous multi-robot specialization for logistics: Instead of a monolithic robot, we decompose the task along morphological lines: a wheeled humanoid for manipulation, a wheeled-legged quadruped for terrain transport. This principle of task-morphology matching has been demonstrated in search-and-rescue robotics but has never been applied to commercial logistics operations. Agentic AI with VLM reasoning in unstructured physical environments: The Binabik Agentic OS does not rely on pre-programmed pick lists or barcode-driven workflows. It uses vision-language models to reason about novel objects and situations in real time, enabling adaptation to parcel types, trolley configurations, and hub layouts never encountered during training. This is a fundamental advance beyond the rule-based or narrow-RL approaches used in current warehouse robotics. First real-world deployment of general-purpose robots in a Swiss railway hub: To our knowledge, no research group or company has deployed humanoid or legged robots in an operational railway logistics environment. This POC creates the first empirical dataset and operational knowledge base for this application domain. The combination of general-purpose robot platforms with an agentic AI layer means the system is not limited to a single task. The same robots and software, with additional skills, can extend to other parts of the logistics chain (parcel sorting, van loading, inter-hub navigation) making this approach fundamentally more versatile and future-proof than single-purpose automation.

Project Idea Metadata

• Project Idea Name: SymbioticFlow: Heterogeneous Multi-Robot Cargo Handling for Railway Hubs
• Date: 4/6/2026 8:17:43 PM
• Administrators:

Project Idea Description

1. Executive Summary

SymbioticFlow introduces a radically new approach to automating railway hub logistics: a heterogeneous multi-robot system where specialized robots collaborate through an agentic AI platform to handle the full cargo flow (from rail wagon to sorting belt) without any infrastructure modifications.

A wheeled humanoid robot (Galaxea R1 Pro) provides AI-driven parcel manipulation at the conveyor interface, while a wheeled-legged quadruped (DEEPRobotics LYNX M20 Pro) autonomously transports trolleys across the unstructured hub terrain. Both robots are coordinated by Binabik AI's Agentic Operating System, which uses vision-language models to reason about tasks in real time. The consortium (Binabik AI + ZHAW) has both robots in hand and the AI software stack operational. This Proof of Concept validates a single robotic cell; the architecture scales to full-hub and multi-hub deployment through straightforward replication.

After this proof of concept, the main limitation stopping wider deployment and scalability will be hardware. However, humanoid robots and mobile manipulators with 20-30kg payload will enter the market next year, while custom robots can be built to meet these specifications. The approach proposed in this project (Agentic OS and software for multi-robot coordination without infrastructure modifications) applies to any mobile robot and mobile manipulation platform. We simply use a wheeled humanoid and quadruped because they are sufficient for the proof of concept in this project, while the solution remains hardware-agnostic.

2. Background and Problem Statement

Every working night, Planzer processes tens of thousands to hundreds of thousands of parcels, unloading rail wagons and feeding heterogeneous cargo onto sorting belts for last-mile distribution. This operation is mostly manual: workers navigate railway tracks and through the trains to pull trolleys from wagons, then unload parcels one by one onto conveyor belts, all under extreme time pressure.

The environment is fundamentally hostile to conventional automation. Railway hubs are rented SBB infrastructure where no permanent modifications are usually permitted. Clearances are tight and the environment is cluttered. Layouts differ significantly between hubs (Zurich, Bern, and others each present different spatial configurations). There are no standardized loading patterns: cargo varies in size, weight, shape, and fragility.

This is why traditional approaches fail here:

• Fixed conveyor extensions require infrastructure changes that SBB constraints prohibit, and cannot adapt to varying hub layouts.
• Standard warehouse AMRs need flat, clean floors and structured environments for localization and robust navigation, while they cannot always traverse rails, gaps, or indoor-outdoor transitions.
• Industrial robotic arms require structured, predictable environments with standardized cargo, almost the opposite of a railway hub.

Meanwhile, Europe faces a shortage of 745,000 logistics workers. The overnight shifts involved are physically demanding, ergonomically hazardous, and increasingly difficult to staff. The need for automation is urgent, but the solution must work within the constraints of these environments, not demand that the environments change.

3. Proposed Solution: The SymbioticFlow Architecture

3.1 Design Principle: The Right Robot for the Right Task

Rather than forcing a single robot type to handle everything poorly, SymbioticFlow decomposes the cargo flow into two specialized roles, each matched to the morphology and capabilities of a specific robot platform, coordinated by an intelligent Physical Agentic AI layer.

3.2 The Dexterous Node: Wheeled Humanoid Robot

The Galaxea R1 Pro is stationed at the conveyor belt interface, where it performs the cognitively demanding task of identifying, grasping, and placing heterogeneous parcels. Its wheels allow it to reach trolleys in different places, compared to a fixed robotic arm with limited workspace dimensions.

Key characteristics include:

• AI-driven perception: Using Vision-Language Models (VLMs), the robot visually assesses each parcel's type, size, orientation, and potential fragility before grasping, enabling zero-shot adaptation to parcel types never seen during training.
• Dual-arm manipulation: Two 6-DoF arms with parallel grippers (26 DoF total) execute precise pick-and-place operations. The wheeled base provides micro-positioning adjustments around the trolley. Vision-Language-Action models are trained to handle the packages.
• Capacity: Rated for 7 kg dual-arm payload (10 kg maximum), covering a large number of parcel weight classes. The 170 cm height and 0–200 cm vertical operating range match the full trolley unloading envelope. The solution is easily deployable to robots with higher payload that will reach the market in the next 6-12 months.
• Onboard compute: NVIDIA Jetson AGX Orin enables real-time VLM inference with end-to-end latency under 200 ms.

3.3 The All-Terrain Tugger: Wheeled-Legged Quadruped

The available LYNX M20 operates as the logistics runner, autonomously transporting trolleys between rail wagons and the sorting station. The robot will be modified with a hitch specific to couple with Planzer's standard trolleys.

• Terrain traversal: Its hybrid wheeled-legged design switches seamlessly between wheeled mode (fast, efficient on flat surfaces) and legged articulation (stepping over rails, navigating gaps, handling uneven terrain). It can handle light slopes and small gaps.
• Trolley towing: Equipped with a mechanical hitch mechanism, the LYNX crouches, slides under the trolley base, locks in, and tows it. Rolling resistance on wheeled trolleys is low: the challenge is the attachment/detachment cycle and terrain navigation, not raw pulling force.
• All-conditions operation: IP66-rated, operational from -20 to 55 degrees C, with bidirectional lighting. If needed, it can be exposed to weather.
• Perception: Dual 96-line LiDAR units (360x90 degree FOV) and wide-angle cameras provide robust SLAM-based navigation through cluttered, dynamic hub environments.

3.4 The Agentic Operating System: Binabik AI's Software Platform

The robots are coordinated by Binabik's three-layer Agentic OS:

1. Cognition Layer: VLM/LLM-based reasoning decomposes high-level goals ("unload wagon 7") into robot-specific task sequences, adapting in real time to unexpected situations (blocked paths, unusual parcels, equipment changes).
2. Execution Layer: A Model Context Protocol (MCP) interface dispatches specialized skills: VLA policies for parcel manipulation, reinforcement learning (RL) policies for hitch engagement, ROS 2/Nav2 modules for navigation.
3. Control Layer: Safety-aware whole-body control ensures compliant, predictable robot behavior in shared human-robot workspaces.

This architecture is inherently modular: new skills can be added for additional tasks (sorting, van loading, inspection) without redesigning the system. The same Agentic OS runs across different robot hardware (already validated on Galaxea, Unitree, and DEEPRobotics platforms), making it a natural building block for broader end-to-end logistics automation.

4. POC Scope and Deployment Path

This project validates a single robotic cell: one Galaxea R1 Pro unloading parcels from one trolley onto a conveyor belt, with one LYNX M20 delivering and retrieving trolleys. The POC measures cycle time, grasp success rate, and navigation reliability in a real Planzer hub environment.

For full-hub deployment, the cell is replicated — 50 to 100 parallel cells coordinated by the fleet management layer of the Agentic OS. Larger quadruped platforms or electric tug attachments can substitute or augment the LYNX for heavier trolleys. Next-generation humanoids (Unitree H2, available at ZHAW) extend the parcel weight range. The POC validates the unit economics and technical feasibility that underpin this scaling path.

5. Objectives

1. O1: Demonstrate autonomous unloading of heterogeneous parcels (up to 7 kg) from a standard Planzer trolley onto a conveyor belt, achieving a cycle time of less than 90 seconds per parcel.
2. O2: Demonstrate autonomous LYNX M20 trolley towing through a representative railway hub section, including rail crossing, narrow passages, and the hitch/unhitch cycle.
3. O3: Validate the Binabik Agentic OS as a real-time coordination layer managing task allocation between the two heterogeneous robots.
4. O4: Characterize the parcel weight and size distribution at a real Planzer hub to quantify the coverage of current robot capabilities and inform next-generation system design.
5. O5: Produce a validated cost model and scaling roadmap for full-hub and multi-hub deployment.

6. Implementation Plan (6 Months)

Phase 1: Lab Integration (Month 1–2)

• Visit Planzer railway hub with robots for initial exploration of capabilities with teleoperation
• Construct mock trolley and conveyor belt interface at ZHAW/Binabik lab
• Deploy Galaxea R1 Pro with initial VLA model for parcel grasping (leveraging Binabik's existing pick-and-place skill library)
• Deploy LYNX M20 for indoor navigation and hitch mechanism testing
• Integrate Agentic OS orchestration: task dispatch, robot status monitoring, error recovery
• Deliverable: Both robots operating individually in lab; initial integrated handoff demonstrated

Phase 2: Capability Development (Month 2–4)

• Fine-tune VLA models on Planzer-representative parcel types (varied sizes, shapes, weights, packaging materials)
• Develop robust VLM-based grasp planning for novel and irregularly shaped parcels
• Optimize LYNX navigation for rail-crossing and uneven terrain (simulation first via Isaac Sim, then physical mockup)
• Refine multi-robot coordination: LYNX signals trolley delivery, Galaxea begins unloading sequence, empty trolley triggers retrieval
• Deliverable: Integrated two-robot demo in lab with measured cycle times per parcel

Phase 3: Planzer Test Environment (Month 4–6)

• Deploy at real Planzer hub (test environment access provided as part of the award)
• Conduct systematic trials: measure parcel throughput, grasp success rate by parcel type, failure mode analysis
• Collect real parcel distribution data (weight, dimensions, fragility categories)
• Demonstrate live operation to Planzer stakeholders
• Deliverable: Validated POC with quantitative performance data in operational environment

Phase 4: Analysis and Dissemination (Month 6)

• Compile results into scaling analysis and cost model
• Present findings at Innovate Mobility co-creation workshop (November 2026)
• Prepare follow-on funding proposals (Innosuisse, Horizon Europe)
• Submit results to leading robotics venues (ICRA, IROS)
• Deliverable: Final report, scaling roadmap, academic publication draft, demo video

7. Novelty and Advancement of the State of the Art

Current logistics automation operates at two extremes: highly structured warehouse systems that require purpose-built facilities, and experimental humanoid robots that remain confined to controlled laboratory or factory settings. Neither approach works in the unstructured, constrained, and variable environments of railway logistics hubs.

SymbioticFlow advances the state of the art in three ways:

1. Heterogeneous multi-robot specialization for logistics: Instead of a monolithic robot, we decompose the task along morphological lines: a wheeled humanoid for manipulation, a wheeled-legged quadruped for terrain transport. This principle of task-morphology matching has been demonstrated in search-and-rescue robotics but has never been applied to commercial logistics operations.

2. Agentic AI with VLM reasoning in unstructured physical environments: The Binabik Agentic OS does not rely on pre-programmed pick lists or barcode-driven workflows. It uses vision-language models to reason about novel objects and situations in real time, enabling adaptation to parcel types, trolley configurations, and hub layouts never encountered during training. This is a fundamental advance beyond the rule-based or narrow-RL approaches used in current warehouse robotics.

3. First real-world deployment of general-purpose robots in a Swiss railway hub: To our knowledge, no research group or company has deployed humanoid or legged robots in an operational railway logistics environment. This POC creates the first empirical dataset and operational knowledge base for this application domain.

The combination of general-purpose robot platforms with an agentic AI layer means the system is not limited to a single task. The same robots and software, with additional skills, can extend to other parts of the logistics chain (parcel sorting, van loading, inter-hub navigation) making this approach fundamentally more versatile and future-proof than single-purpose automation.