- Universal Robots’ Gen 7 platform with native Inbolt 3D vision cuts robot setup from weeks to hours, removing the need for jigs, fixtures and external vision hardware.
- ABB Robotics’ Flexley Stack F712 autonomous forklift uses Visual SLAM as its primary navigation method, with ABB reporting roughly 20% faster commissioning than 2D SLAM systems.
- A study published in Engineering in November 2025 recorded a 92.5% navigation success rate for language-guided mobile robots in simulation, with physical-environment performance still unconfirmed.
Universal Robots’ Gen 7 cobot platform, unveiled at IMTS this week, ships with native support for Inbolt’s 3D vision, cutting deployment from weeks to hours and removing the jigs and fixtures manufacturers have relied on for decades. That integration is a useful marker of where industrial robotics stands right now: AI vision research that was confined to labs two years ago is on the factory floor, changing what autonomous systems can perceive and act on.
Cobot Deployment Gets Simpler
The Gen 7 controller’s open architecture lets vision providers build directly on Universal Robots hardware, eliminating external vision PCs and the cabling that came with them. According to Rudy Cohen, CEO of Inbolt, the result is a setup process that no longer requires specialised integration work for each deployment. Partners including Cognex and Schunk contributed to the platform’s launch. The practical payoff is real-time adaptation to variable production environments, which matters most in assembly and material handling lines where part positioning shifts constantly.
Foundation Models for Manipulation
Vision-Language-Action (VLA) models are moving industrial robotics away from pre-programmed instruction sets toward systems that generalise across tasks. Rather than encoding every motion explicitly, VLA architectures let robots interpret visual input and language context together to determine what action to take next, which is why they are increasingly the architecture of choice for manipulation tasks that vary by part geometry or production run. How widely they have been adopted across new deployments in 2026 has not been independently confirmed.
Visual SLAM Changes AMR Navigation
In July 2026, ABB Robotics launched the Flexley Stack F712 autonomous forklift, its first AMR product to use AI-powered Visual Simultaneous Localisation and Mapping (Visual SLAM) as its primary navigation method. Where traditional 2D LiDAR builds a flat map of static surroundings, Visual SLAM builds a 3D model that distinguishes fixed environmental features from transient objects such as passing workers, enabling context-aware pathfinding in dynamic spaces. ABB reports the technology cuts commissioning time by roughly 20% compared to 2D SLAM systems. RGo Robotics’ Perception Engine takes a similar approach, using vSLAM with learning algorithms so AMRs can refine their maps during operation rather than requiring a dedicated setup phase.
AI Vision on the Quality Line
Quality inspection has historically depended on human eyes or rule-based machine vision, both of which struggle with subtle defects and require extensive recalibration between product runs. Modern AI vision handles that differently: systems trained on large image datasets can flag micro-cracks, surface scratches, misalignments and contamination in real time, without a dedicated calibration step for each new part geometry. The gains matter most in high-precision assembly, where consistent part positioning is non-negotiable and rework is expensive. Continuous automated monitoring also reduces reliance on manual spot-checks, which are by definition intermittent and subject to fatigue.
Physical AI in Unstructured Environments
The harder problem is not vision accuracy in controlled settings but generalisation to environments that were never mapped or anticipated. Structured factory floors, with fixed lighting and predictable layouts, have always been easier ground for robotics. Logistics and fulfilment environments are messier: variable lighting, shifting stock, workers moving unpredictably. The architecture shift from programmed instruction sets to systems that observe and respond is what makes unstructured deployment viable, though the gap between lab performance and production reliability remains the friction point for most teams.
Natural Language Meets Mobile Robots
A study published in Engineering in November 2025 by Tian Wang et al. introduced a system that lets mobile robots accept and execute natural language instructions in manufacturing settings. The architecture combines 3D scene reconstruction, VLM-based semantic segmentation and LLM-driven spatial goal navigation, so a worker can describe a task in plain language rather than programming a waypoint sequence. Tests in the AI Habitat simulator recorded an average success rate of 92.5% across navigation subgoals. The simulator result is encouraging, though performance in physical manufacturing environments, where lighting, clutter and sensor noise compound, remains unconfirmed. Xiaomi’s recent work on high-performance robotics on consumer GPUs points toward the infrastructure trade-offs that follow once language-guided navigation moves off the simulator.
The Gen 7 launch is the sharpest example of that pipeline closing: a 3D vision capability that once required weeks of bespoke integration now ships as a first-party feature, setup time measured in hours.
