Company
Our Approach
Kynetic Intelligence is built on a simple thesis: robot intelligence should not be tied to a single body.
Today’s robots are one-offs. A controller built for a warehouse arm doesn’t work on a humanoid. A humanoid policy doesn’t transfer to a quadruped. Every new robot requires a new stack. This is the bottleneck.
Our architecture separates task reasoning from physical execution. A high-level policy decides what to do. A low-level controller handles how to move. Between them is a learned embedding space — a shared language of movement and intent. This means:
- Embodiment-agnostic high-level reasoning. One policy. Multiple robots.
- Grounded low-level control. Real physics. Contact-rich dynamics. Real transfer.
- Composable skills. Behaviors learned in simulation combine into novel sequences.
We’re methodology-agnostic — not committed to any specific neural network architecture, training algorithm, simulator, or hardware platform. The architecture is the bet.
Training Pipeline
Our three-stage pipeline is designed for data efficiency and sim-to-real transfer:
- Pre-training — Diverse simulated tasks across embodiments build broad capabilities
- Supervised fine-tuning — Human demonstration data via accessible consumer hardware
- RL fine-tuning — Task-specific reinforcement learning on target objectives
Founder
Miguel Alonso Jr. is the founder and CEO of Kynetic Intelligence.
- PhD in Electrical and Computer Engineering, NSF Graduate Research Fellow
- Former lead of ML-Agents at Unity Technologies
- Visiting Associate Professor at Florida International University
- $1M+ in secured research funding
- Developed full-body teleoperation and sim-to-real systems for humanoid and bi-manual robots
Status
We are in the simulation phase — building infrastructure, training pipelines, and evaluation protocols. Hardware is on a 12-month horizon. Our core research question: can a hierarchical architecture with a learned embedding interface achieve better sim-to-real transfer than direct action prediction?