Addverb targets warehouse automation gap with wheeled humanoid Elixis-W

As humanoid robots move from demonstrations to cautious pilot deployments globally, Indian automation firms face a harder version of the same question: how do you build a machine capable enough to be useful and simple enough to be commercially viable in a market where labour remains cheap?

When Addverb Technologies unveiled a wheeled humanoid at LogiMAT India in Mumbai earlier this year, the choice of form factor was itself a statement about where the technology actually stands. The machine, a robot torso mounted on an autonomous mobile robot base, was not the company's most advanced model. A bipedal, fully walking version exists. They chose not to launch it.

That decision reflects a tension running through the global humanoid industry. The robots that attract attention walk and run in front of the cameras. The ones that generate revenue stay upright, keep moving, and don't break after thousands of warehouse cycles.

The wheels-versus-legs debate

Most industrial environments do not need a machine that can climb stairs. Warehouses are already optimised for wheeled movement. Certification is simpler. And wheeled locomotion is significantly more energy-efficient than bipedal walking, which is important given that most humanoids today operate for roughly two hours before needing a recharge, according to Bain's 2025 technology research. Bir Singh, co-founder and CSO of Addverb, said the priority ordering was clear: customer environment first, then safety and stability, then route efficiency, then cost. "Industrial customers value predictability over spectacle," he said.

Where the performance gap sits

The Elixis-W is currently achieving 180 to 220 picks per hour in pilot environments, against a target of 250 to 300. Singh is direct that the gap is not mechanical. It lies in perception, confidence, contextual understanding, and decision-making in semi-structured, unpredictable environments. In early testing, the system recorded depth miscalculations on overlapping stock-keeping units, gripper over-rotation during recovery, and minor collisions on retraction. The engineering response shifted the philosophy from "decide and execute" to "decide, validate, and adapt" — adding redundant depth fusion, pre-grasp confidence scoring, and mid-motion validation pauses.

How the AI stack is structured

Simulation handles gross motor coordination, navigation, and collision avoidance. Real factory floors refine fine manipulation and responses to variable lighting and dynamic obstacles. Safety sits entirely outside the learning loop. Speed limits near workers, torque thresholds, emergency stops, and geofenced zones are all hard-coded. Every perception output carries a confidence score; if it drops below a threshold, the robot slows, re-scans, or hands the task to a human. "The system is engineered to acknowledge doubt rather than mask it," Singh said. The critical failure mode in industrial autonomy is not a robot that stops — it is one that acts wrongly with high confidence.

The data bottleneck

Physical AI systems cannot be trained on internet-scale data the way language models can. Teaching a robot to pick an oddly shaped object from a cluttered bin requires purpose-built visual datasets generated through recorded human repetition. Singh argues India's large workforce gives it a cost advantage in producing this data at scale, a potential edge over well-capitalised Western competitors, though one yet to be demonstrated at commercial scale.

The India ROI problem

The global humanoid market is projected to grow from 2.9 billion dollars in 2025 to 15.3 billion dollars by 2030, according to MarketsandMarkets. India's share depends on whether buyers commit before the ROI case is fully proven. In a market with low labour costs, the standard cost-displacement argument for automation rarely closes. Singh's view is that early adoption will be driven by hazard mitigation, gas-risk inspection, sewer maintenance, and sustained heavy lifting, rather than headcount reduction. Quick commerce is accelerating that case, pushing item-level pick complexity into a territory where human labour is high-volume but physically costly, and where existing automation performs poorly.

The component dependency remains the structural constraint. High-precision actuators, harmonic drives, and specialised chips still require imports. Whether India can close that gap depends on aligning its software capability with hardware precision and global certification standards, a combination that, as Singh put it, is still "a long road ahead."