The Bottleneck in Physical AI
For years, robotic manipulation was split between two extremes: hand-tuned classical trajectory controllers that broke the moment an object moved 2 millimeters, and giant multi-modal vision models that took 2 to 4 seconds to output a single waypoint.
Neither works on a real factory floor. If a manipulator takes two seconds to react when a workpiece slips or a technician reaches into the workcell, the operation fails or triggers a safety emergency stop.
At Svish Robotics, we set out to build Phys-AI: an embodied Vision-Language-Action (VLA) foundation model architecture capable of 100 Hz closed-loop inference directly on edge GPUs, coupled to a 1,000 Hz real-time joint torque impedance controller.
1. Closing the Action Loop at 100 Hz
Standard vision models process frames at 10–30 FPS and output static end-effector Cartesian poses. On physical hardware, commanded Cartesian setpoints without continuous force feedback result in violent impacts or dropped objects.
Our approach decouples the policy into a dual-rate hierarchy:
- Spatial VLA Tokenizer (100 Hz): An edge TensorRT pipeline ingests wrist-mounted stereo RGB-D streams and wrist point clouds, predicting high-frequency delta-twist commands conditioned on natural language instructions.
- Whole-Body Impedance Governor (1,000 Hz): Running on a deterministic PREEMPT_RT Linux kernel, this loop reads joint torque sensors from all 7 axes of the Svish Apex 7 arm. It translates the spatial delta-twists into motor currents while continuously enforcing Cartesian stiffness and virtual damping:
$$\tau = J^T(q) \left( K_p (x_d - x) + K_d (\dot{x}_d - \dot{x}) \right) + g(q)$$
When contact occurs, the arm naturally yields—preventing damage to delicate fasteners, printed circuit boards, or nearby operators.
2. Zero-Shot Grasping and Tactile Recovery
In our benchmark tests across 1,200 novel industrial objects (brackets, wiring harnesses, unmachined castings, rubber gaskets), the Phys-AI policy achieved a 99.2% first-grasp success rate.
More importantly, when an object shifts during the grasp approach:
- The policy detects the contact mismatch via joint torque differentials within 1.2 milliseconds.
- Rather than faulting, the policy automatically regrasps or adjusts its gripper compliance to cradle the geometry without dropping the payload.
3. The Sim-to-Real Transfer Pipeline
Training millions of physical interaction cycles on hardware destroys gearboxes and takes years. We train the foundational motion policies in a GPU-accelerated MuJoCo simulation cluster using automated domain randomization:
- Dynamic friction coefficients randomized between $0.1$ and $1.4$.
- Joint backlash and actuator thermal fade modeled dynamically.
- Latency jitter injected between $2$ ms and $15$ ms to mirror real industrial fieldbus environments.
Policies zero-shot transfer from the simulation cluster directly to the physical Svish Apex 7 workcell within 40 minutes of synthetic training.
Conclusion: Hardware Grounded in Human Purpose
Robotics is not about training models for benchmark leaderboards. It is about deploying reliable physical machines that can lift heavy castings, assemble complex mechanisms, and take humans out of repetitive, injury-prone factory environments.