For Embodied AI, Direct Drive Says the First Bottleneck Is Getting There

A robot can have strong perception and sophisticated planning yet still fail at the first practical requirement of deployment: reaching the job site. Direct Drive Tech’s thesis is that physical access—not intelligence alone—sets an early limit on embodied AI. If a robot cannot cross a curb, threshold, broken pavement section, slope or muddy patch, it cannot perform work there or collect the real-world interaction data needed to improve its behavior. 3437

That argument does not mean software is unimportant. Embodied AI requires a physical body, environmental interaction and intelligence working together, and the body’s capabilities constrain how far the system can operate. 19 Other industry analyses also identify scarce real-world data, weak generalization, hardware maturity, compute and reliability as interconnected barriers. 41 Direct Drive’s distinctive claim is about sequencing: solve the robot’s ability to enter diverse environments first, then expand its intelligence and business coverage.

The physical-access bottleneck

Most indoor mobile robots are designed around relatively predictable surfaces. Their operating assumptions become harder to maintain when a route includes a raised threshold, a curb, stairs, loose soil, damaged pavement, a steep transition or a change between indoor and outdoor space. A robot that stops at those boundaries may still work well inside a structured facility, but it cannot cover the full task or gather data from the places where its mobility fails.

Direct Drive and related reporting describe this as a carrier problem: the missing ingredient is not simply more recording capacity or more training budget, but a platform capable of reaching difficult conditions. A flat-floor robot cannot generate movement data for stairs or mud while remaining on a flat floor. Simulation and video can help with pretraining, but research on embodied intelligence also points to a persistent gap between controlled or simulated environments and dynamic physical interaction. 3445

This leads to a different deployment philosophy. Instead of rebuilding the environment around a robot with dedicated ramps, visual markers, restricted lanes or other infrastructure, the robot should adapt to ordinary human-built spaces. That is especially important for applications that move across sidewalks, buildings, service areas and outdoor terrain rather than operating inside one carefully engineered zone.

Why combine wheels and legs?

The wheel-versus-leg debate is often presented as a choice between efficiency and adaptability. Direct Drive’s approach is to combine them in one mobility system.

• Wheels are well suited to fast, efficient travel over smooth floors and roads, but their performance is limited by step height, gaps and uneven or loose surfaces.
• Legs can raise the body, change posture and negotiate discontinuities, but legged movement typically brings greater control, energy and engineering demands for routine travel.
• Wheel-legs use rolling as the default mode and reserve articulation for lifting, stepping, stabilizing and adjusting the body when terrain demands it.

NVIDIA describes TITA in similar terms: the robot combines the speed and agility of wheeled mobility with the adaptability of legs, using direct-drive joints and hub-motor drive technology. 1 Direct Drive’s broader proposition is therefore not that every route should be walked, but that a robot should be able to roll efficiently until the environment requires a more capable movement strategy.

In practical terms, that could mean rolling along a corridor, using its leg joints to negotiate a threshold, adjusting its posture on a slope and then returning to wheeled travel outdoors. The design is aimed at reducing the number of places where a mission must stop or a site must be modified.

What direct-drive actuators add

The mobility concept depends on the robot being able to make quick, controlled changes in posture and force. Direct-drive wheel-hub motors deliver propulsion at the wheel, while direct-drive or quasi-direct-drive joint architectures provide torque for articulation and obstacle negotiation. TITA is specifically described as using direct-drive joints together with hub motors. 17

Reducing reliance on conventional transmissions can support more direct torque control and responsive movement. It does not eliminate engineering trade-offs: the motors, structure, thermal system and control software must still handle high loads and repeated impacts. Direct drive is therefore not a substitute for intelligence. It is the physical layer that determines whether a planned action can be executed reliably on irregular ground.

TITA illustrates the integration of these layers. Direct Drive’s specifications list eight degrees of freedom, a peak torque of 120 N·m, a 10-kilogram moving payload, hot-swappable batteries, a Jetson Orin NX 16 GB processor and quoted 100 TOPS AI performance. The platform also supports ROS 2 and broader software and hardware interfaces for development. 91132

Its sensing package includes binocular cameras, ultrasonic sensing and SPAD sensors, giving the robot the perception hardware needed to detect and respond to obstacles. 9 The important product idea is not any one specification in isolation, but the connection between sensing, decision-making and a body capable of changing its position when a route is no longer flat.

From Xingtian to TITA and D1

Direct Drive’s product sequence shows a progression from demonstrating wheel-legged mobility to building more configurable platforms.

Xingtian: an early commercial wheel-legged proof point

The company presents Xingtian as an early commercial direct-drive two-wheel-leg robot. Later company and industry materials describe it as combining wheeled speed with legged obstacle-crossing ability and supporting development through an open SDK. 21 Its significance in the product story is that it established the basic proposition: a compact wheel-legged body can target terrain that would challenge a conventional wheeled chassis without committing to fully legged locomotion everywhere.

TITA: a developer-oriented platform

TITA expanded the concept into an eight-degree-of-freedom wheeled-biped platform. It adds a more capable actuator and sensor system, modular mounting rails and interfaces intended for research and application development. The published specifications include ROS 2 support, hot-plug batteries, active obstacle avoidance, a 20-centimeter maximum jump height and a maximum speed listed at 3 m/s, with higher speeds requiring API access. 91132

The platform is positioned for inspection, delivery, logistics, agriculture and research use. Those are company or product-positioning claims rather than proof that every application is commercially solved, but they show the intended path: use one mobility base across tasks that span structured facilities and less predictable outdoor routes. 321

D1: changing the robot’s configuration

D1, also called D-Infinite in some coverage, pushes the idea beyond a fixed wheel-legged body. The platform can operate as a standing two-wheel-leg unit or combine modules into a lower four-wheel crawling configuration. 26

The cited specifications claim a maximum obstacle height of 70 centimeters and payload capacity of up to 100 kilograms in crawling mode; another listed configuration supports an 80-kilogram standing payload. 24 The product’s modular structure is intended to let the robot trade compactness and maneuverability for stability and load-carrying capacity depending on the task.

That matters for mixed environments. A standing configuration can be advantageous where the robot needs a narrower profile or greater access around sidewalks and stairs. A lower, joined configuration can provide a more stable base for heavier loads and rougher terrain. The underlying idea is that general-purpose capability may come not from one perfect morphology, but from a platform that can change its physical arrangement as the job changes.

What this strategy does—and does not—prove

Direct Drive’s architecture addresses a genuine constraint: a robot that cannot reach a setting cannot work there or collect interaction data there. Its wheel-legged approach offers a plausible middle ground between the efficiency of pure wheels and the terrain adaptability of pure legs, while direct-drive actuators connect high-level movement plans to responsive physical control. 143

But physical entry is a first gate, not the entire bottleneck. A robot still needs reliable perception, manipulation, navigation, safety behavior, energy management, affordable maintenance and models that generalize across changing conditions. Government and industry analysis also highlights the scarcity and poor reusability of high-quality physical-world data, the gap between laboratory and real environments, and hardware reliability at scale. 41

The commercial evidence should likewise be read carefully. Direct Drive’s product pages and company materials describe capabilities, applications and performance under stated configurations; those claims are not equivalent to independent proof of general-purpose autonomy or large-scale deployment. Independent reporting on the company’s robotics business has described commercialization as still limited, despite growth in its related product lines. 36

The larger lesson for embodied AI

Direct Drive’s proposed order of operations is straightforward:

1. Build a body that can enter more of the environments where work occurs.
2. Use that access to execute tasks and collect richer real-world interaction data.
3. Improve perception, control and generalization through repeated deployment.
4. Expand from specific applications toward broader business coverage.

That is why the company’s wheel-legged robots are more than an alternative chassis design. They represent a view about where the embodied-AI stack should begin. Before a robot can be genuinely general-purpose, it must first have the physical freedom to encounter the world it is supposed to understand.