Dyna Robotics unveiled the Dyna 2.1 robot on 30 September 2026 and built the pitch around a single metric: mean time between human interventions. The machine is meant to work a full commercial shift, not perform one task in front of a camera.
Key takeaways
- Dyna 2.1 is a semi-humanoid: a torso with two arms on a wheeled base with four steerable wheels
- The arms take parallel-jaw grippers or dexterous hands
- Its world action model was trained on a million hours of human and robot data
- The maker claims task learning in hours, deployment in minutes and 24/7 operation
- Target sectors are hotels, laundromats, food service and data centres
The metric is intervention, not grasp success
The company optimises for MTBI — mean time between interventions, the average stretch before a human has to step in. That inverts the usual narrative: instead of advertising the success rate of a single motion, Dyna measures how long you can forget the robot exists.
The consequence is an emphasis on self-recovery. The robot is supposed to pick up a dropped item and retry a failed grasp without asking an operator. Over a shift measured in hours, those small things decide whether supervision costs more than it saves.
Claimed deployment figures
| Application | Throughput | Accuracy / quality |
|---|---|---|
| Warehouse and fulfilment | 200 units per hour | ±2 mm placement |
| Hotel and housekeeping | 22 minutes per room | ≥97% first-pass inspection |
| Food service | 100 plates per hour | under 40 seconds per plate |
These are vendor figures, without independent verification and without any breakdown of how many of these deployments are in production versus pilot.
The stack: orchestrator on top, policy underneath
Dyna's architecture combines a workflow orchestrator, a vision-language model, the DYNA-2 policy, a world action model and a whole-body controller. The policy was trained with reinforcement learning in simulation, and perception fuses vision, touch and proprioception.
The split is practical: the vision-language model breaks an order into steps, the policy executes the motion, the whole-body controller coordinates driving with reaching. Commercial laundry is a fair test here, because a towel has no fixed shape and every grasp looks different.
Why it matters
In embodied AI the era of single-skill demos is ending. Two companies in two days — Nucleus first, now Dyna — have put a supervision figure front and centre. The weight is shifting from "can the robot do it" to "what does watching it for eight hours cost". Only the second number settles the economics of a deployment.
What's next?
- Dyna has not published an MTBI value in hours or a robots-per-operator ratio, so the 24/7 claim stays unverifiable
- The deployment results come from the vendor, with no indication of which are production and which are pilots
- The chosen sectors — hotels, laundries, food service — are high-turnover ones, marking the real market: where labour is short, not where precision is prized
Sources
- DYNA Robotics — DYNA-2.1
- The AI Insider — Dyna Robotics Launches New Robot for Full-Shift Work
