Figure’s robots are distancing themselves from Joe Biden.
The Silicon Valley startup just shared new footage of a fleet of its Figure 02 humanoid robots showing off their improved human-like walking gait. On the X social media platform, founder Brett Adcock said he was tired of seeing comments in Figure’s replies comparing its humanoid’s bipedal locomotion capabilities to the former president’s.
“We’ve had enough of this,” he said in the post. “Our primary focus has been shipping robots commercially, so this wasn’t a top priority. However, we now have a neural net capable of walking naturally like a human.”
We've had enough of this!
— Brett Adcock (@adcock_brett) March 25, 2025
Our primary focus has been shipping robots commercially, so this wasn't a top priority
However, we now have a neural net capable of walking naturally like a human 💅 pic.twitter.com/acTbpm7I8N
“Walk Like Biden” became a meme in the summer of 2024 following the release of a satirical adaptation of The Bangles’ 1986 hit, “Walk Like an Egyptian.” The parody music video features a montage of Biden’s public appearances, emphasizing moments when his gait appears hesitant or unsteady.
The meme permeated discussions about robotics, with social media users drawing parallels between Biden’s walking gait with rising humanoid robots like the Figure 02, Tesla’s Optimus, and XPENG’s Iron.
Figure’s new demo shows 10 humanoid robots controlled by a single neural network without any adjustments. The fleet of robots walk in perfect sync using the same AI brain. The zero-shot deployment was made possible with NVIDIA technology that simulates years of walking practice in just a few hours.
During simulation training, digital twins of the robots trained to handle trips, shoves, friction changes, and other possible disruptions before attempting anything in the real world. The robots adjust their walking with torque feedback sensors, checking their movements a thousand times per second.
The high-speed feedback loop helps the robots get around similar to humans, whose muscles and nerves are constantly changing during movement. The reinforcement learning approach awards the virtual replicas of the robots points for copying human-like movements, staying balanced, and using energy efficiently. Once training is finished, the learned behaviors are transferred to the physical robots with a process called Sim2Real, short for Simulation to Reality.
Figure is seen as one of the most formidable challengers to Tesla, which aims to manufacture 5,000 Optimus humanoid robots in 2025. The startup recently unveiled its BotQ robot factory, which is intended to scale to 12,000 humanoid robots per year.
The company is facing stiffening competition from the likes of 1X Technologies, Agility Robotics, Apptronik, Boston Dynamnics, and a growing number of Chinese firms that also rely on reinforcement learning to achieve human-like movements.
Shenzhen startup EngineAI has been especially disruptive with its frequent demos showcasing its robots’ increasingly natural and acrobatic capabilities. Engine developed custom actuators and a novel neural network that continuously adapts movements. The company was founded in 2023 and generated serious buzz when it emerged in fall 2024 with its SE01 humanoid’s shockingly lifelike walking gait.
Hangzhou-based Unitree Robotics combines imitation and reinforcement learning. Using the dual approach, Unitree’s humanoid and quadrupedal robots mimic human movements and refine their actions through trial and error. Unitree offers open-source resources supporting reinforcement learning implementations on their robots.
Meanwhile 1X Technologies, which operates out of California and Norway, is collaborating closely with NVIDIA to bring its NEO humanoid robots into homes at scale.
The startup shared footage of its NEO Gamma robot washing dishes autonomously after being trained on a custom dataset collected from real homes. The system uses NVIDIA’s Groot N1 neural network in a 5Hz vision action loop. They trained the robot at a 1X employee’s real home.
Like Figure, 1X Technologies plans to start testing its AI robots in real homes this year.
Another perceived leader in the emerging market, Canadian startup Sanctuary AI, is also leveraging a simulation learning approach as it focuses on robotic hand dexterity. The company recently shared a video clip showing a Sim to Real transfer of a robotic hand trained for manipulation tasks.
With NVIDIA’s Isaac Lab and physics simulation engine, Sanctuary says it can train its proprietary robotic hands across thousands of simultaneous simulations. The Vancouver-based startup recently showed off the robotic hand’s integration with the firm’s flagship humanoid named Phoenix.
With 21 degrees of freedom (DoF), the hand’s design uses miniaturized hydraulic valves for stronger power density compared to fully electric systems. The valves have been tested to withstand 2 billion cycles without leakage or degradation.