How does Chinese startup Direct Drive argue that the main bottleneck for embodied AI and general purpose robots is not intelligence but physical access to real world environments, and how do its wheel legged designs and direct drive actuators address the limitations of pure wheel

Direct Drive’s thesis is that a robot’s first commercial problem is not merely “thinking,” but getting its body into the places where work and data collection occur. If it cannot cross the physical discontinuities of ordinary environments, better perception or foundation models cannot create the real-world interaction data—or business coverage—needed for a general-purpose robot. 67

The “physical entry” argument

• Ordinary environments contain curbs, thresholds, broken pavement, slopes, stairs, muddy ground and transitions between indoor and outdoor areas. The company’s argument is that conventional wheeled robots often cannot reach such sites, so the associated physical-interaction data cannot be collected in the first place. 6
• This reframes the constraint as a carrier/body problem rather than solely an AI-model problem: simulation has limits, and a robot excluded from a setting cannot learn from, or provide service in, that setting. 67
• Consequently, Direct Drive argues against deployment models that make the environment conform to the robot—such as dedicated ramps, markers, narrowly restricted lanes, or otherwise engineered routes. Its aim is for the platform to adapt to human-built environments instead.
• The claim is a strategic proposition, not an established industry consensus: embodied AI also depends on manipulation, safety, perception, compute, data quality, cost, and reliability. Still, the company’s emphasis is consistent with the broader definition of embodied AI as a system that senses, decides and physically affects its environment. 3

Why wheel-legs rather than wheels or legs alone

• Pure wheels are efficient, fast, energy-conscious and practical on smooth floors and roads, but are constrained by step height, gaps, loose surfaces and uneven terrain. 6
• Pure legs can negotiate obstacles, but generally impose greater energy, control and cost burdens for routine long-distance travel; Direct Drive’s materials characterize them as less efficient and less scalable than wheeled mobility. 5
• Wheel-legs seek to retain rolling as the default movement mode, then use articulated legs to lift, step, stabilize, adjust body posture and recover when terrain requires it. NVIDIA’s description of TITA similarly says it combines wheeled speed and agility with legged adaptability through direct-drive joints and hub-motor drive technology. 4
• The practical intent is to traverse curbs, uneven pavement and slopes without stopping the job or redesigning the route, while providing a base that can operate across mixed indoor–outdoor spaces. This explains the company’s view that household robots may enter homes first by “rolling,” rather than by adopting a fully humanoid gait.

Role of direct-drive actuators

• Direct-drive wheel-hub motors provide propulsion directly at the wheel, while direct-drive or quasi-direct-drive joint motors supply the high-torque, fast-response articulation needed for posture changes, stepping and obstacle crossing. 4
• Eliminating or reducing conventional transmission components is intended to improve torque control and instantaneous motion response; the trade-off is that motor, thermal, control and structural design must be strong enough to manage the resulting loads.
• In this architecture, AI selects and plans movement, but the direct-drive hardware is what makes the selected movement physically executable on irregular terrain. That is the core of Direct Drive’s claim that embodiment cannot be solved by software alone.

Platform evolution

• Xingtian / Xing Tian (2021): Direct Drive’s first commercial direct-drive two-wheel-leg robot, positioned as the early proof that a wheel-legged body could offer more terrain access than a conventional wheeled chassis. 7
• TITA (2023): An eight-degree-of-freedom wheeled-biped platform, adding a more capable direct-drive joint system and a development-oriented hardware/software architecture. 5
  • Direct Drive says TITA has eight quasi-direct-drive actuators, 120 N·m peak torque, eight degrees of freedom, hot-swappable batteries, open Linux-kernel/API access and ROS 2 support. 5
  • Its sensor/AI configuration includes an NVIDIA Jetson Orin NX 16 GB, quoted 100 TOPS compute, ultrasonic sensing, SPAD sensing and binocular cameras; upper-body rails support modular accessories. 45
  • The platform is designed for inspection, logistics, agriculture and research, with the manufacturer claiming active obstacle avoidance, posture adjustment, drop resistance and self-recovery. 45
• D1 (2025): A modular embodied-robot platform that can combine robot units and switch between two-wheel-leg and four-wheel crawling/quadruped-style configurations. 5
  • Direct Drive’s cited specifications claim a maximum 70 cm obstacle height, up to 100 kg payload in crawling mode, and a configurable multi-unit architecture rather than a fixed single morphology. 5
  • The significance is not just extra payload: a single configuration can favor compact urban movement and stair/sidewalk access, while a joined, lower four-wheel crawling arrangement can favor stability, load carriage and rougher terrain.

How this supports its broader business claim

• The company presents its deployments and component shipments as evidence that direct-drive mobility is not only a laboratory concept. Its corporate materials cite large-scale global use of its technology, although deployment figures and “world-first” labels are largely company or prospectus claims and should be treated accordingly. 57
• Its sequence—Xingtian to TITA to modular D1—shows a progression from proving wheel-leg locomotion, to providing a sensor-rich and developer-accessible platform, to adapting physical form and payload capacity to the task. 57
• Thus the proposed order of operations is: first secure physical entry into diverse environments; then accumulate real interaction data and reliable task execution; then broaden applications; and only after that approach a genuinely general-purpose robot. 67