General Trajectory

The most interesting thing about General Trajectory is the shape of its published work, which is more specific and less sweeping than most physical-AI pitches. Rather than claiming one model controls everything, the company surfaces three dated research releases with numbers attached.IronBench is the standout. It frames transformer design as reinforcement learning with verifiable rewards and reports 74.4% compliance across 900 held-out specifications. Verifiable rewards matter here: the reward signal is checkable, which is a different and generally more credible setup than preference-based training on motion. It also signals that GT is not only about robots — scientific and engineering design work sits under the same 'physical world' umbrella, alongside the stated application areas of advanced manufacturing and scientific R&D.Phoenix (March 2026) is an autonomous counter-UAS system: AI drone interceptors with fully onboard terminal guidance, described as resistant to jamming. Onboard guidance is the detail that matters — it implies the interceptor does not depend on a live uplink to a ground controller, which is exactly the failure mode jamming exploits. This is defense work, and it puts GT in a different procurement and regulatory context than a factory automation vendor.The dexterous manipulation update from November 2025 uses reward-guided imitation learning — preference-guided learning, per the site — and reports 63% gains on hard objects. Hard-object manipulation is where grip policies usually degrade, so it is a sensible place to publish a delta.What is missing is equally important. The site's own navigation is team, about, updates, contact. For a buyer, that means the evaluation is a conversation with a research team, scoped to one task and one machine, not a platform procurement. Performance numbers are the company's own and are not independently benchmarked. Safety-critical certification is not established. If your line needs deterministic microsecond-latency control, a learned policy is the wrong tool regardless of vendor.Where it fits: a research or advanced-engineering team with a concrete, measurable objective — a manipulation task, an inspection or intercept problem, a design-search problem like transformer specification — and the appetite to co-develop. Where it does not: a production line, a hobbyist, or anyone who needs to sign up and run something today. Hebbian Robotics' hflow, launched on Hacker News in August 2026, is a useful compare point if your bottleneck is robotics data plumbing rather than policy learning.