Two modes
Solo is full local execution on one robot with no network dependency. Crew is every robot on the site network sharing one governed memory and acting as a team. Solo is a first-class state, not a fallback: a robot that loses the link keeps working with a smaller planner and conservative traffic rules. Crew adds intelligence over the network and never takes over the control loop, the safety functions or the manipulation policy.
What runs where
- Servo loops and whole-body control — On the robot — 500 to 1,000 Hz
- Safety monitors: e-stop, geofence, speed and separation — On the robot — 100 Hz, under 10 ms
- Manipulation policy, 3B-class VLA — On the robot — 10 to 25 Hz
- Perception and scene graph — On the robot — 1 to 10 Hz
- Local planner, 7 to 8B model — On the robot — 40 to 80 tokens per second
- Episodic capture — On the robot — 1 Hz
- Fleet arbitration through Open-RMF — Site server — 100 ms to 1 s
- Site memory service — Site server — Under 1 s reads
- Task planner and verifier, 30 to 70B model — Site server — 0.5 to 3 s per call
- Skill library, nightly fine-tunes — Site server — Hours
Why each one sits where it does:
- Servo loops: hard real time, never on a radio.
- Safety monitors: the standards assume local safety. Measured 5G e-stop paths show tails of 100 to 150 ms.
- Manipulation policy: measured on Jetson Thor. A server policy is faster in the lab and unsafe on radio tails.
- Perception and scene graph: needed in Solo.
- Local planner: one reasoning step every few seconds. Anything larger competes with the policy for Thor's memory bus.
- Episodic capture: captions with pose and time into a local index, rolling 30 to 60 minutes.
- Fleet arbitration: traffic, lifts, doors and task allocation.
- Site memory service: multi-writer reconciliation needs one authority.
- Task planner and verifier: called at task boundaries or on exception. Cannot fit beside the policy on the robot.
- Skill library: checkpoints pushed with version pins, never mid-task.
Three memories
Core memory is the site's identity, small enough to pin into every planner's context, and changed only through a human review gate. Short-term memory is the live shared world state that every robot writes to and replicates locally. Long-term memory is the episodic history with provenance, the knowledge distilled from it and the skill versions, consolidated on the site server. Robot proposals move into core memory only when a named person approves them, which is the same crew-plus-human-gate pattern the rest of Tidewell runs on.
Four failure modes are designed against from the start: leakage across scopes, stale propagation, contradiction persistence and provenance collapse. We will publish a small benchmark for multi-writer robot memory alongside our own results.
The contract
Link-loss detection under two seconds. Automatic downgrade to Solo with no human action. Append-only re-sync on reconnect; the robot never blocks on upload. No policy and no safety function runs off the robot. No checkpoint is pushed mid-task. The numbers sit in the Tidewell Brain specification table and turn from design targets into measured values as tests complete.
Where we start
We do not train a base vision-language-action model from scratch. NVIDIA's GR00T is released under a commercial licence, Physical Intelligence's π models are open, and Alibaba, Ant Group and others give away planners and policies that run on Jetson-class compute. The base layer is a commodity. Our budget goes to post-training on data from our customers' sites, to the memory and coordination layer, and to security.
The last mile
Task-specific post-training on deployment data is what crosses the economic threshold. Our target for any task we sell is 95 percent success without intervention, and we publish the number. Corrections captured on site feed the next training run and, through the human gate, the fleet's memory.
What we do not claim
No manipulation policy runs off the robot over wireless. No weights update across the fleet in real time; fleet learning is minutes to nightly. No 70-billion-parameter model runs on the robot. No whole shift is held in the policy's context; memory is external and retrieved. No cross-site global brain. No Huawei Ascend anywhere in the stack.
Compute
Jetson Thor T5000 on the biped, T4000 on the wheeled manipulator with an Orin cost-down variant, one carrier board for all three. Training runs on our own GPUs in Singapore. We keep a clean capitalisation table, with no Chinese ultimate parent and no controlling Chinese strategic investor, so that access to training hardware and site servers is never in question. We state that plainly to investors.
What we publish
Evaluation results, including failures, in the Brain category of Insights. The competitor deep dive behind this page is in Landscape.