Elon Musk knows how to capture attention and create illusions. While other car manufacturers showcase greater driving ranges or more comfortable interiors, he makes a car hang upside down and leap into the air. As an engineer, I find it fascinating to watch, because such a spectacle requires knowledge, experimentation and a skilled team. At the same time, it raises the question of what it actually proves: an extraordinary manoeuvre performed under controlled conditions, or the everyday usability of a car of the future? asks Professor Kristo Vaher, Head of the Robotics Engineering Curriculum at TTK University of Applied Sciences.
At a time when Tesla has discontinued production of two car models in order to begin manufacturing humanoid robots on a large scale, the same question arises. A human-like machine may run, dance and lift objects in a video, but it only becomes valuable to an employer when it can perform a required task safely, independently and repeatedly. What connects the Roadster demonstration and the development of humanoid robots is precisely the gap between a successful demonstration and a reliable product.
A large supply of compressed gas and a few seconds in the air
The Roadster’s compressed-gas thruster system was developed through collaboration between Tesla and Musk’s other company, SpaceX. From an outside observer’s perspective, Musk’s line of thinking might seem simple: if technology suitable for controlling a rocket already exists, why not try attaching it to a car? In 2018, SpaceX had already sent a Roadster into space. First, a car on a rocket; now, rocket technology on a car. As a spectacle, it works brilliantly.
According to Tesla engineer Lars Moravy, the demonstration vehicle contained a total of 600 litres of pressurised gas tanks (roughly equivalent to three bathtubs of water), around 250 kilograms of compressed air and ten thrusters. Releasing the gas at high speed generates a reaction force that can accelerate, lift and rotate the vehicle.
The tanks, piping and valves also add weight. In a simplified calculation, 250 kilograms of gas would be enough to keep a vehicle weighing a couple of tonnes hovering for approximately 6–7 seconds, assuming the entire supply is used and an effective exhaust velocity of 500 metres per second. This is an order-of-magnitude estimate, not a measured flight duration. Take-off, rotation and braking for landing all consume the same supply.
A separate system is used to attach the car upside down to a ramp. Fans extract air from beneath the car, while edge seals restrict the flow of air back underneath. The resulting pressure difference presses the vehicle against the surface. However, the seal in contact with the asphalt wears down, larger gaps increase air leakage, and maintaining the required suction force becomes more difficult. Tesla’s patent describes adjustable skirts, but their actual durability remains to be proven. As an engineer, I would next like to see several consecutive fast laps on a race track, as well as stable performance through corners, under braking and when crossing kerbs.
Tesla plans to use compressed-gas technology in cars that will eventually go on sale. Nevertheless, the large gas supply and a special capability lasting just a few seconds raise questions about practicality. A road car must perform reliably in different weather conditions and after years of use.
At the same time, Musk is preparing a transformation at his car factory. Tesla has allocated space previously used for the Model S and Model X production lines at its Fremont factory to the manufacture of Optimus humanoid robots. This brings the same question of practicality from cars to robots: when is a new machine reliable enough to be entrusted with everyday work?
Real work is already underway
The use of humanoid robots is no longer limited to laboratories and demonstrations. According to logistics company GXO, Agility Robotics’ humanoid robot Digit has handled more than 100,000 packages at the company’s warehouse in Georgia. The robot’s task is to pick up a package from a pallet and place it on a conveyor belt. This is a solution integrated into the company’s actual workflow, with plans to expand its use to other warehouses.
At BMW’s Spartanburg plant, the Figure 02 humanoid robot placed sheet-metal components into a welding fixture. According to published figures, the robot moved more than 90,000 components and accumulated approximately 1,250 operating hours during the pilot project. These examples illustrate the current strong application area for humanoids: repetitive material handling in a prepared workplace. However, there is a difference between independently completing a single work cycle and independently managing an entire working day. Lifting boxes does not mean that a robot can solve every problem that arises in a warehouse or move to another department to replace a colleague. Nor does a high number of completed movements alone reveal how often human intervention was required. In its overview, Figure lists a target of more than 99% successful placements and zero human interventions during a shift, but does not disclose whether these targets have actually been achieved.
A human shape does not automatically mean human abilities
The main purpose of a humanoid robot is to operate in an environment designed for humans. The dimensions of workbenches, shelves, doors and tools are based on the human body. A robot with two arms and an appropriate reach could use some of the same infrastructure and move between different tasks. However, a wheeled robot may be simpler and more energy-efficient on a smooth warehouse floor, while a conventional industrial robot may be faster and cheaper in a fixed workstation.
Imitating human movements is difficult because even picking up a cup requires precise coordination between perception and movement. We use our eyes to assess the cup’s position, feel the contact with our fingers and adjust our grip according to its weight and how slippery it is. A robot must combine information from cameras, touch sensors and force sensors, while controlling its joints quickly and precisely. In a walking robot, picking up an object also affects the body’s balance.
A robot may detect that a drawer will not open, but determining whether it is locked, something is obstructing its movement or its grip is incorrect can be difficult. The solution depends on the cause: should it change the direction of the pull, try gripping again or stop the action? Sensors provide only partial information about a situation, and a previously trained solution may not work in a new context.
A simple calculation illustrates why this matters. If a task consists of 100 consecutive actions, each with a 99% probability of success, the probability of completing the entire sequence without a single error is only around 37%, assuming independent outcomes: 0.99 to the power of 100. This does not necessarily mean the entire task will fail if the robot can correct its own mistakes. However, it explains why a single successful video clip is not proof that a robot can independently complete an entire work shift.
In my assessment, the main limitation is reliable autonomy: the ability to understand a situation, choose an appropriate action, carry it out and recover from mistakes when necessary. Battery capacity also limits operating time. Manufacturer specifications for several well-known models indicate approximately two to five hours, depending on the robot and its workload. Charging and battery replacement must therefore be factored into work schedules. These breaks, however, are a more clearly defined problem than dealing with every possible unexpected situation.
Why is Musk in such a hurry?
Replacing car production lines with humanoid robot manufacturing is a major bet on technology that is still under development. Why isn’t Musk waiting until the robots’ capabilities have been more thoroughly proven?
Musk justifies robot production, among other things, by the need for training data: machines need real-world practice. He has spoken about building tens of thousands of robots to collect data, while acknowledging that artificial intelligence, reliability and service life still need further development.
I see a clear business rationale behind this rush: early production makes it possible to gain experience, identify shortcomings in real-world use and develop manufacturing expertise. Musk is betting that production and technological development can progress in parallel. Waiting also comes at a cost: competitors may gain an advantage in training data, customer relationships and service capabilities.
Tesla is not operating in an empty market. Figure is working with BMW, Agility is gaining experience in GXO warehouses, Boston Dynamics is developing Atlas with Hyundai, and Apptronik is collaborating with Mercedes-Benz and Google DeepMind. China’s UBTECH has also announced deliveries of industrial humanoid robots. Not all these companies are at the same stage of development, but the race to build a useful and commercially viable robot is already underway.
When will humanoid robots arrive in Estonian workplaces?
Meeting a robot during a demonstration and working alongside one every day are two very different experiences. My prediction is that the first pilots involving narrowly defined tasks in Estonian companies are realistic within the coming years. The most likely settings are warehouses and manufacturing facilities, where work areas can be prepared and technical support is close at hand. This is an assessment of the direction of development, not a confirmed deployment schedule.
Wider adoption will ultimately depend on economics. If a robot constantly needs an engineer working alongside it, it may be a valuable development project but an expensive employee. Employers care about actual productivity, downtime, maintenance requirements and the cost of completing a single task. Choosing a solution should start with the question of which task needs to be automated. Only then should we decide whether a human-shaped robot is actually necessary.
A humanoid’s value is measured by the work it does
The purpose of humanoid robots should be to take over physically demanding, dangerous and tedious tasks from people. Industrial robots and collaborative robots have already been performing such tasks successfully for years: welding, painting, operating machine tools, assembling products and stacking heavy packages onto pallets. Humanoids are entering a field where effective solutions already exist.
The opportunity for humanoids lies in taking on tasks that have so far been too difficult, inflexible or expensive to automate. For example, tasks that require moving between different workstations and handling various objects in an environment designed for humans. A humanoid must demonstrate why it is the right solution for these tasks.
In the case of the Roadster, I would like to see repeatable performance on a race track. In the case of a humanoid, I want to see a robot that completes its work, detects mistakes, deals with them and starts the next cycle. And does so every working day. That is when a humanoid begins to become a colleague.
* Opinion artcile was published on November 12, 2026 in Äripäev Online.

