If you want to build a fake HAL 9000, all you need is an LED, some carpentry skills, and any laptop accessing a talking AI. If you want to build your own R2D2, you’ll have a tougher job assembling a range of materials, motors, and electronics. But what if you wanted to make your own working version of TARS, that bizarre, blocky robot from Christopher Nolan’s Interstellar that looks like a stainless steel ATM with metal posts for legs sprouting from its shoulders?
Well, to do all that, you might need to hold a master’s degree from Carnegie Mellon University’s Robotics Institute and be a senior robotics engineer at Nimble.ai. Fortunately for Aditya Sripada, he just happens to be one, which is how he and his longtime collaborator Abhishek Warrier built TARS3D.
They also wrote “Walking, Rolling, and Beyond: First-Principles and RL Locomotion on a TARS-Inspired Robot,” which was a Mike Stilman Award Finalist for outstanding papers at the 24th IEEE RAS Humanoids Conference in Seoul, the Olympics for humanoid robotics research.
As the following demonstration video shows, TARS3D is exciting, with four independently articulated, telescopic “pillars” that immediately and quickly transform (from a side-view) into an X-shape as pillars 1 and 3 rotate forward and pillars 2 and 4 rotate back. TARS3D also extends its curved pads on the tops and bottoms of each pillar as “feet,” all the better for rolling as an eight-spoke double rimless wheel. Sripada and Warrier claim it’s the first TARS-emulating robot that can both walk and roll.
TARS3D robot
Of course, Nolan’s TARS had things easier: while not being a computer-generated image, it was a human-sized puppet whose operators were digitally deleted from the screen, and in wheel-form, was attached to a motorized, amphibious dolly.
While the roboticists note in their paper that much of robotic locomotion research is biomimicry-focused, robots moving through “many human-engineered settings can benefit from nonanthropomorphic forms.” Just like the movie bot, TARS3D can also walk (if a bit tippily), but rolling is definitely its strength. That’s thanks to the robot’s seven independent movements (three rotary and four prismatic) which “use machine learning and optimization to identify gaits not tractable through analytic methods.”
In their paper, the authors describe how they “used deep reinforcement learning (DRL) in simulation” and “observed that the learned policy can recover the analytic gaits under the right priors and discover novel behaviors as well.” They learned that TARS3D’s “biotranscending morphology” led to “multiple previously unexplored locomotion modes,” and that further exploration will open a promising pathway for multimodal robotics.
Aditya Sripada
Although TARS3D isn’t ready for interstellar missions just yet – it’s still cable-connected, and at 25 cm (9.8 in) and 990 g (2.2 lb) of 3D-printed components, it’s small and light enough to stand on a table – eventually, Sripada and Warrier will test how well TARS3D moves across a range of terrains.
Sripada said that building TARS3D – which he began back in November 2022 without a laboratory, funding, or affiliation, “just late nights and weekends and a desire to reconnect with the simple joy of building robots,” – reminded him of why he “fell in love with robotics in the first place … the wonder, the patience, the heartbreak when things fail, the quiet euphoria when they finally work, and the feeling that somewhere in the process, you discover a small new truth about motion, persistence, and yourself.”
New Atlas asked Sripada to discuss his motives for creating TARS3D and how the prototype might lead to further solutions and opportunities. His emailed responses are below.
NEW ATLAS: What aspects of TARS were such an inspiration for you to create TARS3D?
ADITYA SRIPADA: Before the answers, a small context point. TARS3D is primarily about locomotion as a mobility primitive, built to evaluate the practical feasibility of gaits shown in the movie. So, I will center on how it fits within decades of legged research and where that leads.
TARS is a simple rectangular body that appears to do a lot, which led me to a clear question: can a low-complexity robot both roll for transit and step across discrete footholds? Rolling looked cinematic at first, but with small changes to the contact arc and a brief telescopic push it works in practice, and the math matches classical rimless wheel results while the same body walks with stable foot placement. The structure also offers redundancy because many contact edges can carry load, and we also see diverse gaits through deep reinforcement learning.
NEW ATLAS: What are the practical applications of TARS3D for helping human beings at home, in workplaces, at leisure and entertainment, or during emergencies?
SRIPADA: [While this question doesn’t] have direct or immediate answers, I’ve included brief reflections that might help illustrate what directions could be possible.
TARS3D is a mobility module that rolls for transit and steps when the path breaks into sparse footholds. In warehouses, plants, and infrastructure sites, it moves quickly in corridors, then steps over pallet gaps, cable trays, grates, stair edges, and cable trenches to reach sensors, verify inventory, or perform inspection, with mission fixtures and contact tuning added as needed.
NEW ATLAS: Are there ways that TARS could be useful for helping animals including pets, or helping ranchers with their livestock?
SRIPADA: This is secondary to the locomotion result, but the same quiet and low profile motion could support routine checks at gates and water points and short remote observation in clinics or shelters. Any near-term use would layer sensing and handling policy on top of the mobility.
NEW ATLAS: Are there any applications of TARS3D technology for exoskeletons, EVA repairs of vehicles and stations, or extraterrestrial exploration?
SRIPADA: The focus on stable transitions, few joints, and energy-aware motion can guide lighter assistance devices. For work outside a vehicle, a compact module could ride rails in roll mode, then step and brace to hold a sensor or a tool during inspection. These directions follow naturally from the validated locomotion.
