Data AnalysisVerified

Reviewed and published by trentmaziarz, April 25, 2026. Discovered and drafted by our automated research pipeline.

Foundation uses data generated during Phantom robot deployments to retrain its proprietary AI action models, according to CEO Sankaet Pathak in a March 2025 interview. When a human operator remotely intervenes to correct a robot's behavior, that intervention is labeled and fed back into the model. Sacra reports that this feedback loop is built into Foundation's business model, with larger fleets producing more training data and improving model accuracy over time.

Details

According to Sacra, Foundation's business model is structured so that each robot deployment generates training data that feeds back into its proprietary AI models. When a human operator intervenes via teleoperation to correct a robot's behavior, that intervention is labeled and used as training data to improve the model, as described by CEO Sankaet Pathak in a March 2025 interview. This approach means that AI model performance is directly tied to the scale of the robot fleet; a larger deployed fleet produces more diverse operational data, enabling more robust model generalization over time.

Products affected

Phantom MK1

Sources & Evidence

News Article

https://www.humanoidsdaily.com/news/foundation-emerges-with-phantom-humanoid-betting-on-novel-actuators-and-hybrid-ai

News Article

Sankaet Pathak, CEO of Foundation, on why humanoids win in robotics

Sacra·Mar 2025

https://sacra.com/research/sankaet-pathak-foundation-humanoids-robotics/

News Article

Foundation funding, news & analysis

Sacra·Mar 2025

https://sacra.com/c/foundation/

Cite this record

Trace Foundation. (2026). Foundation: Foundation uses data generated during Phantom robot deployments to retrain its proprietary AI action models, according to CEO Sankaet Pathak in a March 2025 interview. When a human operator remotely intervenes to correct a robot's behavior, that intervention is labeled and fed back into the model. Sacra reports that this feedback loop is built into Foundation's business model, with larger fleets producing more training data and improving model accuracy over time (data as of 2026-04-25) [Data set record]. AI Trace. https://www.aitrace.org/r/practice/4bc66879-e1d7-443b-bcea-0c67e6425d8c. Accessed September 10, 2026.

Stable link: https://www.aitrace.org/r/practice/4bc66879-e1d7-443b-bcea-0c67e6425d8c
Data as of: April 25, 2026
Last verified: April 25, 2026

How to cite AI Trace

Other practices by Foundation

OtherFoundation tested two Phantom MK-1 humanoid robots in Ukraine starting February 2026 in a frontline-reconnaissance support role, per co-founder Mike LeBlanc's statements to TIME. The robots use a camera-first visual perception system and LLM-based task-to-motion software that converts operator commands into movement sequences, and they operate with a human in the loop for any lethal decisions.OtherFoundation announced a planned AI system called 'fleet coherence' that would allow multiple Phantom robots to share real-time task and environment data, coordinate their actions across a shared world graph, and dynamically allocate work to avoid duplication. As of early 2025, this system is in early development and robots largely operate independently.ProductivityFoundation deployed the Phantom MK1 humanoid robot, which uses an onboard AI system to autonomously perform repetitive physical tasks — such as moving cases and racking car parts — in manufacturing and logistics facilities. The robot uses camera-based perception and a proprietary physics-action model to navigate environments and execute tasks without continuous human intervention.

Have evidence about Foundation's AI practices? Submit a report.

Report a Sighting →