Tesla AI lead on unified models
Tesla AI lead Ashok Elluswamy said at GTC that self‑driving and humanoid robotics architectures share core elements—joint high/low‑level decisioning in a single model for real‑time safety—with scalability to higher DoF and richer sensors. The comment frames a growing view that unified model approaches can span vehicles and legged/bipedal robots. (x.com)
Ashok Elluswamy is listed by Tesla as vice president of AI software and has taken leadership responsibility across Autopilot (FSD) and the Optimus humanoid program. (indianexpress.com) Tesla publicly describes its roadmap as a single end‑to‑end neural architecture—often framed internally as a “neural world simulator”—that the company says is intended to be shared between vehicle autonomy and humanoid control. (tesla.com) Tesla’s Full Self‑Driving (Supervised) fleet has generated billions of real‑world miles of video and telemetry data—Tesla’s live FSD safety page currently reports multiple billion cumulative miles driven with FSD engaged—data the company says will feed unified model training. (tesla.com) On the compute side, Tesla has invested in custom training and inference stacks (Dojo/D1 and successive Dojo plans) and iterated its in‑car AI4/HW4 compute to support larger vision‑first neural nets used across cars and robots. (applyingai.com) Internally, Tesla shifted Optimus training away from motion‑capture and teleoperation toward a vision‑only, camera‑based data pipeline in mid‑2025, a change reported by Business Insider and industry outlets as aligning robot training with Tesla’s car‑focused data strategy. (africa.businessinsider.com) Elon Musk’s public production targets for Optimus have been aggressive—public comments have ranged from tens of thousands of units in early years to a stated goal of scaling to roughly 500,000–1,000,000 robots annually by 2027—though independent coverage has repeatedly flagged technical and deployment gaps. (wccftech.com) Elluswamy has reiterated the unified‑model thesis across technical forums and social posts—publishing deep‑dive talks at ICCV/ScaledML and amplifying short public comments on X—signaling a sustained, cross‑platform push inside Tesla to merge vehicle and humanoid learning pipelines. (youtube.com)
