Skild AI Releases S1: Foundation Model Evaluated Across Humanoid and Quadruped Morphologies

Cross-Embodiment Policies: Unified Model for Heterogeneous Hardware

On August 31, 2026, Pittsburgh-based robotics foundation lab [Skild AI] (founded by Deepak Pathak and Abhinav Gupta) released [Skild S1], a foundation model evaluated on physical manipulation and locomotion across distinct robot chassis.

Multimodal Pretraining Architecture

Rather than training isolated policies per hardware setup, Skild S1 trains on four data modalities: 1. Teleoperation Trajectories: Joint angles, end-effector poses, and gripper states. 2. Egocentric Human Demonstration Videos: Video demonstrations of tool use and object handling. 3. Physics Simulation: Parallel environmental rollouts for contact dynamics. 4. Sensor Glove Data: High-frequency tactile and finger pressure trajectories.

Benchmark Results

In reported benchmark evaluations:

Unseen Task Success Rate: 66.0% (Zero-shot in-context video prompting) compared to 9.0% for baseline language-only policies.

Task Horizon: Up to 10 minutes of continuous multi-stage execution across tasks including appliance operation and material handling.

Tool Handling: The model performed functional tool substitution (e.g. using a spatula to flip objects) based on visual demonstration without task-specific retraining.

Application to Autonomous Robot Leagues

In autonomous events such as the WHRG Dexterity Challenge and RoboCup Industrial, cross-embodiment models provide initial operational baselines without requiring ground-up reinforcement learning per chassis.