🤖 Embodied AI & robotics
Skill selection, failure recovery, active perception and assistance for robots
Specialty tags
Browse projects by a specific field or use case.Projects in this industry
View all →robo-jev — Physical Judgment Model
An independent 2B model for text-based physical judgments between robot planning and control.
S1 →It sits between task planning and low-level motion control, choosing actions such as continuing, slowing or stopping and judging collision, tipping, spilling or part-seating conditions. Each question uses a fixed prompt and reads answer-position probabilities into choice, noul or score outputs. It does not provide visual perception or replace motor controllers.
Jev Drone
Choose a detour, climb, brake or target reacquisition while code handles flight control.
S1 →In the main experiment, ordinary code converts simulated depth and segmentation buffers into a scene. Jev chooses a maneuver with Choice, assesses risk with Score and judges whether the target is genuinely lost with Noul. Flight control, guidance and safety reflexes remain in code and can override its advice.
Case studies
View all →StartLux Quadruped Emergency Delivery Simulation
Open a door, yield to traffic, move an obstacle, and deliver supplies in a complete simulated task.
StartLux Quadruped 2v2 Football Simulation
One robot advances while another supports: passing, turning, and shooting illustrate team-level tactical choices.
Jev Physical AI — Warehouse Robot Incident Triage
Decide whether a simulated fleet incident needs human help, which team owns it and its urgency.
Jev Robot Control — xArm7 Simulation
Compare bounded robot intent and movement choices in a shared apple-to-plate simulation.
Gemini Robotics ER 2
High-level spatial reasoning, task planning and multi-robot coordination from Google DeepMind.
KnowNo — Robots That Ask for Help
Calibrate candidate actions and ask a person when the next step is ambiguous.