Cybernetic Physics
Cybernetic Physics provides a cloud-based platform that runs robots on real floors by owning both simulation and deployment. The system launches Isaac Sim sessions on cloud GPUs, routes commands from simulation to robot fleets, and enforces safety envelopes during live operation. The company also develops batch-invariant VLA inference techniques to ensure deterministic robot behavior across varying batch sizes.
- Robotics
- Artificial Intelligence
- Industrial Automation
- Software Only
Funding
Funding not disclosed
Founders
Founder details are not available yet.
Product
Problem
Robotics teams train policies extensively in simulation, but transferring that behavior to physical machines remains unreliable. Non-deterministic inference, where identical inputs produce different outputs due to floating-point precision and kernel reduction order changes, can propagate through action spaces and affect real-world robot behavior. Deployment infrastructure that connects simulated training to live operation is fragmented, making it difficult to own the "sim-to-real" gap.
Solution
Cybernetic Physics provides a platform that runs robots on real floors, taking responsibility for both simulation and deployment. Their system launches Isaac Sim sessions on cloud GPUs with browser control, lets agents design simulation environments from prompts, and provides a control plane that routes commands from simulation to robot fleets. The platform supports teleoperation with session recording for training data, policy rollout with canary testing, safety envelope enforcement, and GPU compute rebalancing between training and live inference. Their technical team also publishes methods for achieving batch-invariance in VLA inference, such as tracing numerical divergence and implementing custom Triton kernels for deterministic operation.
Target Audience
Primary customers are robotics companies and research teams deploying fleets of manipulation arms or mobile robots that need reliable sim-to-real transfer, VLA model inference, and continuous operation in production environments.
Features
- Cloud GPU orchestration with session launch, idle timeout, capacity pooling, and rebalancing between training and live inference workloads
- Agent-driven simulation environment design from text prompts, with automated asset resolution and lineage tracking
- Control plane with fleet heartbeats, watchdog alerts, leader election, and policy version enforcement to reject incompatible commands
- Teleoperation infrastructure that records every motion as training signal and supports handoff back to autonomy mid-task without reset
- Safety envelope monitoring with motion clamping, emergency stop testing, and workcell geofence updates
- Batch-invariant VLA inference techniques, including tensor tracing to find first divergence and custom Triton kernels (Conv2d and attention BMM) to eliminate batch-size-dependent reduction order changes
- Real-time VLA inference on AMD hardware via FlashRT port to MI300X, using hipBLASLt, AITER, and custom MFMA kernels
- Python SDK and MCP server for external agent access, with authentication, rate limiting, and audit logging
- Session recording and replay, including video-to-COLMAP reconstruction of real footage back into simulation