Published on

Watch: Foundation Puts Phantom Through a Gauntlet of Marbles and Banana Peels

Humanoids Daily
Written byHumanoids Daily
The Phantom humanoid robot navigates a path of thousands of small, colorful marbles scattered across a black floor mat. The unit is tethered from above for safety during the high-difficulty balance exercise.
The Gauntlet: In what the company describes as the "final test for all the marbles," the Phantom relies on foot torque sensors and internal IMUs to maintain stability on a highly unstable surface without visual guidance.

Foundation Robotics has released a new video titled "The Walking Test," trading its usual militant rhetoric for a more experimental—and occasionally comedic—look at the Phantom MK1’s physical resilience. The footage shows the humanoid navigating a series of household hazards, ranging from Lego bricks and packing peanuts to the "age-old question" of banana peels.

While the presentation is lighthearted, the engineering goal is serious: discovering the "edges of its abilities" to refine the robot's fundamental mobility.

The "Spinal Cord" vs. The Brain

The demonstration highlights a specific subset of the Phantom’s architecture: the "whole-body controller." Described by Foundation as the robot's "spinal cord," this system translates coarse movement commands into fine-tuned adjustments for the six motors in each leg.

Notably, this specific test was conducted "blind." Unlike the six-camera array housed in the Phantom's head assembly, the balance policy used in this video ignores visual data. Instead, it relies on:

  • Foot Torque Sensors: Determining if and how the robot is making contact with the ground.
  • IMU Sensors: A suite of Inertial Measurement Units located in the legs, pelvis, arms, and head to track the "gravity vector" and the robot's center of mass.

By ignoring vision, Foundation is testing the robot's ability to react to physical perturbations—like slipping on a marble—solely through proprioception and haptic feedback. This contrasts with the company’s higher-level Deep Variational Bayes Filters (DVBF), which are designed to help the robot "infer" or predict the physics of its environment before it makes contact.

A charcoal-black Foundation Phantom humanoid robot attempts to walk across a floor covered in yellow banana peels while tethered to a safety gantry. The robot is shown mid-stride in a laboratory setting with its internal actuators and structural rods visible.
Slippery Science: Foundation puts the Phantom’s balance and reactive "spinal cord" to the test to discover the "edges of its abilities" when encountering unexpected slippage.

Sim-to-Real and Manual Control

The video clarifies that this specific balance policy was developed using reinforcement learning (RL). The system is trained in a simulation environment through millions of trials of "success and failure" before being transferred to the physical hardware.

During the "Walking Test," a technician used a standard PlayStation controller to direct the robot's gait. While the balance remains autonomous, the direction and rotation are handled by a human operator—a setup Foundation views as essential for defense applications where a "human in the loop" must make high-level tactical decisions.

The Banana Peel Factor

The results of the gauntlet were mixed, highlighting the "sim-to-real gap" that remains a hurdle for the industry. While the Phantom successfully navigated a field of mouse traps and waded through marbles, the banana peel proved to be its undoing. The robot suffered a significant slip and was only saved from a "trip to the Robot Hospital" by its safety gantry.

This reliance on a tether serves as a reminder of the hardware's current fragility. Foundation has previously struggled with public hardware failures, including a high-profile crash at Newsweek’s headquarters. The use of a gantry in this video suggests that while the "spinal cord" is maturing, the Phantom is not yet ready to navigate unstructured, slippery environments without a literal safety net.

Scaling for "Dangerous Tasks"

Foundation frames these experiments as a necessary step toward its goal of a general-purpose humanoid. CEO Sankaet Pathak argues that mastering these "basic" mobility challenges is what will eventually allow the Phantom to perform tasks that are "too dangerous or repetitive for humans"—ranging from automotive manufacturing to setting up mouse traps or navigating battlefields in War.

The company continues to pursue an aggressive roadmap, aiming to scale production to thousands of units in 2026. Whether the Phantom can move from the controlled environment of the "Walking Test" to the unpredictable reality of a commercial or military deployment—without its tether—remains the central question for the San Francisco startup.

Watch the full demonstration below:

Share this article

Stay Ahead in Humanoid Robotics

Get the latest developments, breakthroughs, and insights in humanoid robotics — delivered straight to your inbox.