Past project · 2025–2026
Strike Robot
Strike Robot builds the software stack for physical AI: SR Platform creates simulation-ready worlds and assets from language, SR Agentic gives robots task-shaped understanding of the real world, and both are tested on humanoids, quadrupeds and arms in simulation and on real hardware.
01 / Building simulation worlds by hand
SR Platform
SR Platform generates physics-valid simulation environments, production-grade 3D assets and training-ready datasets from a single natural-language description. Generated assets pass validation gates for geometry, articulation and physics readiness.
Asset creation
3D geometry from text, or a real object reconstructed from photos, with physically based materials.
Spatial layout generation
Full environments composed from language, with physics constraints built in.
Simulation and rollouts
Robot policy rollouts in generated scenes, with sensor, collision and success data exported.
Real-time editing
Swap assets and change physics or lighting without leaving the platform.
02 / Robots acting on stale or hallucinated world models
SR Agentic
SR Agentic is an intelligence layer between perception and action. It turns a natural-language task into a lean, task-shaped world model, runs a fast safety loop on the robot and deep reasoning in the cloud only when something new happens, and verifies before it commits.
Task-conditioned scene graph
Builds only the spatial structure the task needs: the map is a function of the task.
Open-vocabulary perception
Detects and grounds objects it was never trained on, fusing detection, segmentation and captioning.
Real-time adaptation
Change detection updates the world model and the planner re-plans on the fly.
Trust and anti-hallucination
Confidence gating, grounded retrieval and visual checks: the robot abstains and re-checks instead of guessing.
03 / Safety in factories without people
SafeGuard ASF
SR Agentic on the Unitree G1 humanoid for autonomous industrial safety: fire and smoke, abnormal pipeline temperature and intruders in restricted zones, with ReAct reasoning over 23 tools and learned locomotion policies.
04 / Behaviors that work in simulation but not on the robot
Simulation to reality
A real laboratory is reconstructed inside SR Platform so that behaviors validated in simulation can be checked against the real space. The simulation stack is open source.
05 / Autonomy without relying on the cloud
Legged robots
Humanoids and quadrupeds that turn camera, depth and LiDAR data into maps and actions, with on-device inference on Qualcomm edge hardware so they keep working offline.
06 / Physics that is hard to simulate
Contact-rich and deformable objects
Assets with realistic contact, friction and deformation: cloth that can be grabbed, pulled and draped, articulated objects, and physical properties validated in simulation.
07 / From instructions to closed-loop manipulation
Manipulation
Robot arms trained and tested on articulated, physics-enabled assets generated by SR Platform, from opening a microwave to semantic pick-and-place driven by SR Agentic.
08 / Long tasks that combine moving and handling
Mobile manipulation
Legs, wheels and arms working together on multi-step tasks inside environments generated entirely by SR Platform.
Videos: Strike Robot; each clip links to its original post. Descriptions follow the posts and strikerobot.ai.