Fleet brains, embodied memory and on-robot compute, September 2026

Three marks run through this post and mean what they meant in the research file. [single source] is a claim found in exactly one place. [inference] is our own reasoning from the evidence rather than a published claim. [prior knowledge] is something not re-verified in this pass. Everything else has at least two sources or an authoritative primary source in the Sources list.

The pass ran on 5 September 2026: about 60 web searches and 45 page reads across arXiv, NVIDIA documentation and forums, DeepMind and vendor pages, and the trade press. The search budget ran out partway through the compute and supply-chain parts, so Wi-Fi 6E roaming measurements, Intel and AMD edge parts, Qualcomm module pricing and Physical Intelligence fleet work are thinner than intended and are flagged where they appear.

The four parts are: multi-robot shared memory and coordination; memory architectures for embodied agents; on-robot and site-server compute; and the supply-chain constraints a Singapore buyer works under.

Part A. Multi-robot shared memory and coordination

Who shares what, and does it run for real

System (owner, date)	What is shared across robots	How (topology)	Latency class	Real deployment evidence	Notes
Gemini Robotics ER 2 and Gemini Robotics 2 (Google DeepMind, 30 Jul 2026)	Task-level semantic understanding and sub-task handoff, exchanged as natural language between robots. No shared weights, no shared world state, no shared memory store. Each robot runs its own copy of the model stack.	Cloud VLM (Gemini API and Live API bidirectional streaming); VLAs and navigation APIs declared as "tools". Coordination is peer-to-peer over language.	"Sub-second" per ER 2 call; cloud round trip. ER 2 claims 91.3% moment-finding accuracy, 0.96 s mean absolute distance, "4x faster than competing models".	Demos: Apptronik Apollo 2 plus Franka FR3 Duo tool-kitting handoff; a Spot fetch demo. ER 1.6 has been live in Boston Dynamics Orbit AIVI-Learning for all enrolled Spot customers since 8 Apr 2026. On-Device 2 is trusted-tester only.	ER 2 is the orchestrator brain; the on-device VLA (Gemma-based) cannot do multi-robot coordination on its own and is "not evaluated on mobile platforms or whole-body control".
Amazon DeepFleet (Amazon Robotics, arXiv 2508.08574, Jul and Aug 2025)	Fleet movement prediction: robot positions, goals, interactions. Effectively a shared learned model of floor traffic.	Centralized: one batched forward pass over every robot on the floor; "weights are shared across all N agents". Four architectures: Robot-Centric 97M params on about 5M robot-hours; Robot-Floor 840M; Image-Floor 900M; Graph-Floor 13M.	Not published; the forecasting horizon is seconds to minutes, not control-loop.	Trained on data from "hundreds of thousands of robots"; Amazon claims a 10% fleet travel-time improvement at 1M-robot scale.	Homogeneous drive units on a grid. Not a memory system; a traffic oracle.
Agility Arc (Agility Robotics, Mar 2024, updated 2025 to 2026)	Facility map, workflow definitions, KPIs (uptime, throughput, MTBI), robot status; WMS, WES and MES API.	Cloud SaaS. No evidence of shared episodic or semantic memory or shared skills across Digits.	Dashboard and mission latency, seconds.	Commercial Digit fleets (GXO, Schaeffler and others); integrates with MiR and Zebra AMRs; Digit gen 6 moving to Jetson Thor.	A fleet ops tool, not a fleet brain.
Boston Dynamics Orbit (2023 onward; AIVI on Gemini since Apr 2026)	Missions, maps, schedules, inspection findings, 360-degree Site View history, anomaly detection results. Multi-site dashboards.	AWS-hosted, or an on-prem "Site Hub" 1U appliance, or an OVA VM. Robots connect via Wi-Fi or LTE.	Mission scheduling; inspection-review latency, seconds to minutes.	Hundreds of Spot fleets in industry; Stretch supported; Atlas planned. Atlas 2026 units go to Hyundai RMAC and Google DeepMind; the first external Atlas pilot is in 2027.	The claim that "once a single Atlas learns a new task, that task can immediately be replicated across the fleet" is marketing; the mechanism, a policy checkpoint push, is not documented. [single source]
UBTech BrainNet 2.0 and Co-Agent (UBTech, Mar 2025 ZEEKR announcement; Walker S2 in 2025 to 2026)	A cloud "swarm brain" holds fleet scheduling and a scenario database; robots share environmental data, so one robot maps a blocked aisle and others reroute. Co-Agent runs locally: a multimodal LLM plus small skill models plus error handling. An "Internet of Humanoids" central hub.	Hub and spoke: a cloud super-brain plus an on-robot sub-brain; a 5G factory network at ZEEKR Shenzhen.	No latency numbers published.	"Dozens" of Walker S1 at ZEEKR (sorting, handling, assembly, inspection); Walker S2 has a 3-minute autonomous battery swap; Walker S2 units at a China-Vietnam customs post (Jul 2026).	The closest public analogue to a connected brain. The architecture claims are press-release level and independent verification is absent.
NVIDIA Isaac, OSMO and Mega (2025 to 2026)	OSMO is an open-source orchestrator for data generation, training, reinforcement learning and software-in-the-loop pipelines across on-prem and cloud. The Mega Omniverse Blueprint is digital-twin fleet simulation for routing and task optimisation (KION trial).	Cloud or on-prem orchestration of workloads, not a runtime robot memory.	Offline.	Mega trials with KION; OSMO used in NVIDIA's own pipelines.	NVIDIA has no shipped runtime fleet-memory product. Its fleet story is "train in sim, push a new brain".
Open-RMF and RoMi-H (Open Robotics plus Singapore CHART, IHiS, HopeTechnik, GovTech; 2019 onward)	A traffic schedule database of every fleet's intended itineraries, lift, door and passageway state, and task dispatch. Conflicts are resolved by fleet-adapter negotiation with a judge.	A central RMF core plus per-vendor fleet adapters; the MiR adapter is public.	Path-planning timescales, seconds; no hard real time.	Singapore hospital pilots under RoMi-H; RMF is the de facto Singapore public-sector standard for mixed-vendor hospital fleets. [prior knowledge for current hospital coverage]	Shares intent and infrastructure state, not perception or memory.
InOrbit (OpenRobOps), Formant, Viam (2024 to 2026)	Telemetry, teleoperation, missions, over-the-air updates, alerts, data pipelines. InOrbit open-sourced the OpenRobOps data model; Viam does modular hardware abstraction and fleet OTA (Tech Mahindra partnership 18 Jun 2026); Formant is about USD 250/robot/month enterprise.	Cloud SaaS with an edge agent.	Seconds.	Many AMR fleets.	RobOps, not brains. Useful as the ops plane under a fleet brain.
Figure Helix-02 (May 2026)	Nothing explicit: two humanoids run the same VLA policy and infer partner intent "from motion alone", with no message passing and no central planner; all inference is onboard.	Peer implicit.	On-robot VLA rate.	A bedroom reset demo; BMW line work with "8-hour autonomous shifts".	Shows that shared-weights-only coordination works for co-located pairs; it does not scale to site-wide memory.
Learning While Deploying (Shanghai Innovation Institute and AgiBot, arXiv 2605.00416, May 2026)	Policy weights. 16 AgiBot G1 dual-arm robots upload rollouts and human interventions; a central learner (DIVL plus QAM RL) broadcasts a new shared policy every 50 training steps; the policy runs on-robot at 30 Hz.	A central learner with on-robot inference.	Policy refresh every few minutes; 4 h wall clock is about 60 robot-hours.	A lab and industrial testbed, 8 tasks, 95% success.	The best 2026 evidence that fleet learning is real at 16-robot scale. The critic stays on the server.
Fleet-DAgger (UC Berkeley, 2022)	A shared policy learned from interventions by a pool of human supervisors; the ROHE metric; 8.8 times ROHE over baselines.	Central.	Offline updates.	4 ABB YuMi arms, 2 remote humans; a 100-robot sim.	The academic ancestor of every "fleet learns from teleop" pitch.
Robot Fleet Learning via Policy Merging (2023) and the Federated Learning for Cloud Robotic Manipulation survey (arXiv 2507.17903, Jul 2025)	Weights merged from heterogeneous datasets without centralising data.	Peer or hub.	Offline.	Sim plus small real.	Federated learning is still a survey-and-prototype topic for manipulation.
Multi-robot 3D scene graphs: Hydra-Multi (MIT, 2023), MR-COGraphs (Tsinghua, Dec 2024 and Mar 2025), CURB-OSG (Mar 2025), Language-Grounded Hierarchical Planning with Multi-Robot 3DSG (Jun 2025), D-Lite	A fused metric-semantic 3D scene graph of rooms, places, objects and open-vocabulary features.	Hydra-Multi: a centralized server fuses incremental inputs and solves relative transforms plus loop closures. MR-COGraphs: a learned encoder compresses object-feature graphs before transmission, more than 80% data reduction, merged by feature-based place recognition.	Online; the Hydra family runs in real time. No hard numbers for fusion latency.	Real multi-robot experiments, indoor and outdoor at large scale.	This is the technical base for a shared core memory.
LLM multi-robot planners: RoCo (Columbia 2023), SMART-LLM (2023), DART-LLM (2024, DAG task decomposition), EMOS (ICLR 2025, a "robot resume" from URDF, Habitat-MAS), COHERENT (ICRA 2025, Proposal-Execution-Feedback-Adjustment), DynaHMRC (Jun 2026, decentralised), LLM-HBT (Oct 2025, behaviour trees), survey arXiv 2502.03814	A task graph and sub-task allocation; robots' capability descriptions; dialogue transcripts.	Mostly one central LLM assigner plus per-robot executor agents; RoCo and DynaHMRC are dialogue-based and decentralised.	Seconds per planning round on a cloud LLM.	Sim benchmarks; a few tabletop or small-team real demos. No industrial deployment.	Reusable ideas: the robot resume, the DAG task graph, the feedback loop.
Governed Shared Memory for Multi-Agent LLM Systems (arXiv 2606.24535, Jun 2026; MemClaw)	A shared, multi-tenant memory service for agent fleets, with scoped retrieval, temporal supersession, provenance tracking and policy-governed propagation.	A central REST service with write-time enrichment.	Sub-second provenance reconstruction; 97.5% intra-fleet visibility and zero cross-fleet leakage in tests.	Software agents only, no robots.	The clearest statement of the consistency problem, set out below.

What the field actually shares today

Four families of sharing exist in shipped or near-shipped systems, and none of them is a shared brain:

Shared weights, offline. Fleet learning (Fleet-DAgger, policy merging, AgiBot's Learning While Deploying) pushes a new policy checkpoint to every robot. Sharing happens on the timescale of minutes to nightly. At 16 robots this is proven, and the inference still runs on each robot at 30 Hz. Figure's Helix-02 shows the degenerate case: the same weights, zero communication, and coordination that emerges from watching each other.
Shared intent and traffic state. Open-RMF, DeepFleet and every AMR fleet manager share where robots intend to go, plus lift and door state. DeepFleet is the largest learned version, at 97M to 900M parameters over millions of robot-hours with centralised batched inference. This is the only family with hard evidence of fleet-scale economic impact, in Amazon's 10% travel-time claim.
Shared ops data. Orbit, Arc, InOrbit, Formant and Viam share missions, maps, telemetry, KPIs and over-the-air updates. Orbit is the most advanced on AI, with Gemini-backed inspection reasoning since April 2026, but it still shares inspection results, not robot memory.
Shared language. Gemini Robotics ER 2 (July 2026) is the first big-lab product to demonstrate heterogeneous robots handing off sub-tasks. The mechanism is explicit: each robot runs its own copy of the same stack and they talk in natural language; there is no common memory store, no shared scene graph and no shared episodic history. UBTech's BrainNet 2.0 claims more, including a cloud scenario database and rerouting from another robot's blocked-aisle observation, but the claims are press-release grade.

What nobody has shipped is a site-level memory service where heterogeneous robots, a wheeled manipulator and a biped, write to and read from one governed store of a fused 3D scene graph with object permanence, an episodic log with provenance, semantic site knowledge and SOPs, and a skills library, with consistency rules, and where each robot keeps a local replica so it degrades gracefully offline. Every component exists in a paper or a product; the integration, and especially the consistency and provenance layer for embodied writes, does not exist as a product. The academic literature has only just started to name the problem, in Governed Shared Memory (June 2026) and the MemAgents workshop at ICLR 2026.

Part B. Memory architectures for embodied agents

The four memory types

Memory type	Function	Representative implementations	Storage substrate	Can it run on-robot, Thor-class?	Multi-writer consistency issue
Working / short-term	Current scene, object states, task state, last N steps	Hydra (real-time 3DSG, MIT), ConceptGraphs (open-vocab 3DSG from 2D foundation models plus multi-view association), Clio (task-driven open-set 3DSG), REACT (updatable 3DSG), KARMA short-term (object position and state deltas with a replacement policy), HiMe "Sentry" tier (ICML 2026), RoboMemory temporal FIFO with LLM consolidation	A hierarchical graph of rooms, places and objects with per-node embeddings; a FIFO buffer	Yes. Hydra runs in real time on laptop-class CPU and GPU; ConceptGraphs is close to offline but Thor has headroom. It must be on-robot for Solo mode.	Two robots seeing the same object from different frames create duplicate nodes; this needs place recognition plus a relative transform (Hydra-Multi) and node-merge rules.
Episodic	What happened, where and when; retrieval by time, space or semantics	ReMEmbR (NVIDIA: VILA captions plus timestamp and pose into Milvus; memory built on a Jetson Orin 32GB, queried by GPT-4o; Nova Carter, 25-min memory), Embodied-RAG (a semantic forest over a topological map; km-scale, 250 queries), STaR (task-conditioned retrieval, Feb 2026), RoboMemory episodic (RAG extractor and updater, Qwen3-Embedding), eMEM (SQLite plus hnswlib plus R-tree, in-process, 80.8/100 on eMEM-Bench), E-mem (multi-agent episodic reconstruction, Jan 2026)	Vector DB rows of embedding, pose, time and caption; or graph nodes	Memory building yes, with VLM captioning at about 1 Hz on Orin or Thor; long-horizon querying over hours or days is a server job, since FindingDory shows VLMs choke past a few hundred images.	Ordering and clock skew across robots; duplicate events; stale propagation, where robot A's stale observation overrides robot B's fresher one.
Semantic / long-term	Site knowledge, SOPs, object affordances, people, preferences	KARMA long-term (a 3D scene graph as long-term memory; 1.3 to 2.3 times success in AI2-THOR), RoboMemory spatial KG with LLM conflict detection, Mem0 (LLM extract then ADD, UPDATE, DELETE or NOOP; a graph variant), Zep (a temporal knowledge graph; 63.8% against Mem0's 49.0% on LongMemEval), A-MEM (Zettelkasten-style linked notes), MemoryBank (Ebbinghaus decay), Letta and MemGPT (a core in-context block plus recall and archival), HiMem (Jan 2026), BrainMem (Apr 2026), HoloAgent-0 (3D spatial memory, Jun 2026), Embodied Agents Meet Personalization (May 2025)	A knowledge graph plus a vector index; a small "core" block pinned in the LLM context	A read replica on-robot, since the site KG is small at tens of MB. Writes and consolidation belong on the site server.	Contradiction persistence, where two robots assert different states of the same door or shelf; and provenance collapse, where it is lost who said this and when.
Procedural	Skills, policies, code, behaviour trees	Voyager skill library (executable code), Isaac GR00T fine-tuned action heads per embodiment, LLM-HBT behaviour trees, SayCan-style affordance scoring, RoboMemory low-level executor (pi0 with LoRA)	Weights per embodiment, code snippets, behaviour-tree XML, plus an index of when each skill applies	Weights must be on-robot. The skill index and library live on the server with a local cache.	Version skew, where robots sit on different checkpoints; this needs skill versioning and a rollout policy of the Viam or Arc kind.
Benchmarks	Measure the above	FindingDory (60 Habitat tasks; VLMs struggle past a few hundred images), WorldLines (Jun 2026), EvoMemBench, AMA-Bench, MemCompiler, EmbodiedBench (RoboMemory plus 26.5% average success rate with Qwen2.5-VL-72B), Habitat-MAS (EMOS), RoCoBench	not applicable	not applicable	No benchmark yet measures multi-robot write conflicts on a shared embodied memory. [inference]

How the leading systems store things

ReMEmbR (NVIDIA) captions short video segments with VILA, embeds them, and stores embedding, x-y-yaw and t in MilvusDB; an LLM agent does iterative retrieval by text, position or time. Memory building ran on a Jetson Orin 32GB and reasoning used GPT-4o. This is the reference design for a robot remembering where it saw something twenty minutes ago. Its weaknesses, such as confusing a soda machine and a water fountain, come from caption-only memory without a scene graph.
Embodied-RAG (CMU) is a semantic forest over a topological map with language descriptions at multiple abstraction levels; it handles object-level and "ambiance" queries across km-scale environments and is used as a global planner for drones, a LoCoBot and a quadruped.
RoboMemory (Aug 2025) has four parallel modules. Spatial is a dynamic knowledge graph with LLM-based conflict detection; temporal is a FIFO of step summaries with consolidation; episodic and semantic are RAG with an extractor, an updater and a vector store using Qwen3-Embedding. There is a planner with a critic. The backbone is Qwen2.5-VL-72B and the executor is pi0 with LoRA on Mobile ALOHA. The real-world lifelong-learning result is that success improves on the second attempt at 15 tasks. This is the closest single paper to a full local memory stack, but the 72B backbone cannot live on the robot.
KARMA (Sep 2024) makes long-term memory a full 3D scene graph and short-term memory object position and state changes with an adaptive replacement policy; it reports 1.3 and 2.3 times success and 3.4 and 62.7 times efficiency gains in AI2-THOR.
HiMe (ICML 2026) is an explicit answer to the frequency-competence paradox: a high-frequency Executor (the VLA), a Sentry (working memory that watches for events) and a Planner (a slow, big model), with add, update and delete memory plasticity. It reports 90% average success on long-horizon manipulation.
eMEM (Jun 2026) is a hybrid graph memory with three indexes (SQLite for structure, hnswlib for semantic, R-tree for spatial), tiered "hippocampal to neocortical" consolidation and ten LLM tool-callable recall primitives, running in-process with no infrastructure. In its ablation a plain RAG baseline loses 30 points on context-dependent retrieval and 29 on false-memory rejection.
Mem0, Letta, A-MEM and Zep come from the software-agent world, and the practical lesson is a small pinned core memory block in context, LLM-driven extract-then-reconcile (add, update, delete, no-op) at write time, and a temporal knowledge graph that beats flat vectors for "what was true when" (Zep 63.8% against Mem0 49.0% on LongMemEval). Robotics has not yet adopted the last two systematically; RoboMemory's KG conflict detection is the closest.
Governed Shared Memory, or MemClaw (Jun 2026), names the four fleet-memory failure modes: unauthorized leakage, stale propagation, contradiction persistence and provenance collapse. It proposes scoped retrieval, temporal supersession, provenance tracking and policy-governed propagation, with write-time synchronous enrichment. It covers software agents only, but it transfers directly.

The on-robot and site-server split for memory

This split is our own reading of the evidence above [inference].

Layer	On-robot (Thor T5000 or T4000)	Site server	Sync rule
Scene graph (working)	Built and updated locally at sensor rate; about 1,000 to 10,000 nodes per zone	A fused site graph, Hydra-Multi style	The robot pushes node deltas, compressed MR-COGraphs style at more than 80% smaller, at 1 to 5 Hz when connected, and pulls the zone graph on entry
Episodic	A rolling buffer of captions and poses for the last 30 to 60 min, a few MB; a local vector index	Full history, days to months; Milvus or pgvector; consolidation jobs	Append-only upload; the server assigns global order; the robot never blocks on upload
Semantic (site KG, SOPs, people)	A read replica of tens of MB, with a pinned core memory block of a few KB in the planner context	The source of truth; LLM extract-reconcile on write; provenance and supersession	Pull on change; robot writes are proposals until the server reconciles
Procedural	Policy weights, skill index cache, behaviour trees	Skill library, versions, rollout policy, fine-tuning	Over-the-air with a per-robot version pin; never mid-task
Fleet arbitration	Local fallback: an RMF-style itinerary with conservative rules	RMF traffic schedule and negotiation; task DAG allocation	If the link is lost for more than 2 s, the robot switches to Solo-mode rules: slow, yield everything

The consistency problem, concretely

When several robots write to one memory you get, in order of pain: duplicate objects, where the same pallet seen by two robots becomes two nodes; stale overwrite, where robot A reports a door open at t=100 and robot B reported it closed at t=130 but A's message arrives later; contradiction persistence, where both survive in the knowledge graph and the planner picks either; and provenance collapse, where an SOP learned from a wrong human correction propagates fleet-wide.

Solutions that already exist and are cheap to adopt: CRDT-style key-value "virtual stigmergy" for cell-level belief maps; last-writer-wins with hybrid logical clocks for object state; write-time LLM reconciliation, as in Mem0 and RoboMemory; scoped visibility and provenance metadata, as in MemClaw; and Hydra-Multi's rule that geometry is only merged after a loop-closure-quality relative transform is found. The hard, unsolved part is semantic contradictions, such as "this corridor is off-limits during ward rounds", where the fix is human-gated promotion.

Part C. On-robot compute reality

The NVIDIA Jetson Thor family

Primary sources: the NVIDIA product page, JetPack r38 docs, the CNX Software price table of 22 Jul 2026 and the ProventusNova roadmap.

Module	AI compute (FP4 sparse)	Memory	CPU	Power	Price (1KU) before and after 22 Jul 2026	Availability
Jetson T5000	2,070 TFLOPS; 2,560 CUDA and 96 Tensor cores Blackwell up to 1.57 GHz	128 GB 256-bit LPDDR5X, 273 GB/s	14-core Arm Neoverse-V3AE, 2.6 GHz	40 to 130 W (the default NVP mode is 120 W; MAXN unconstrained)	USD 2,999 then USD 4,999	Shipping since Aug 2025
Jetson T4000	1,200 TFLOPS	64 GB LPDDR5X, 273 GB/s	12-core Neoverse-V3AE	40 to 70 W	USD 1,999 then USD 2,999	Shipping
Jetson T3000	865 TFLOPS	32 GB, 273 GB/s	8-core	"about half of T5000"	not disclosed	Q1 2027, emulation in JetPack 7.2.1 [single source]
Jetson T2000	400 TFLOPS	16 GB	not published	not published	not disclosed	Q1 2027 [single source]
AGX Thor Developer Kit (T5000)	as T5000	128 GB	as T5000	130 W	USD 3,499 then USD 5,499	Out of stock on the NVIDIA US marketplace in Jul 2026; available via distributors
AGX Orin 64 GB (fallback)	275 TOPS INT8 sparse	64 GB LPDDR5, 204.8 GB/s	12-core Cortex-A78AE	15 to 60 W	USD 1,599 then USD 2,999	Available to Jan 2032
AGX Orin 32 GB	200 TOPS	32 GB	8-core	15 to 40 W	USD 899 then USD 1,799	Available
Orin NX 16 GB	100 TOPS	16 GB	8-core	10 to 25 W	USD 599 then USD 999	Available

Key points:

The July 2026 price increase, 50 to 101% across the Jetson line with no press release and discovered via NVIDIA's FAQ and marketplace, coincides with the DRAM squeeze: LPDDR5X contract prices rose 89% quarter on quarter in Q2 2026, and lead times reached 26 to 39 weeks for mobile LPDDR5X and 40 to 58 weeks for some advanced memory. NVIDIA end-of-lifed every LPDDR4 Jetson (Xavier, TX2) with last orders on 1 Jul 2026 and last ship on 15 Jul 2027, citing DRAM market dynamics. Module lead times of 6 to 9 months and non-cancellable, non-returnable terms on volume orders follow from these data points [inference].
The ECCN for Jetson modules and dev kits is 5A992.c (NVIDIA FAQ). The product lifecycle is 5 to 10 years of availability with a 3-year warranty on production modules; JetPack 7 (Ubuntu 24.04) supports Thor now and Orin during 2026.
Thermal: the module TDP is 130 W, and a junction-to-package thermal resistance of 0.231 K/W implies roughly a 30 degC junction-to-module delta at full load; the passive throttle trip is near 110 degC and critical shutdown near 115 degC (JetPack r38.2.1 docs). Forum threads asking NVIDIA for a power breakdown for enclosed robots went unanswered beyond "read the Thermal Design Guide". The practical consequence is that a battery biped budgets the 70 W of a T4000 or a 60 to 80 W capped T5000 profile, not 130 W [inference].
Memory bandwidth is the real limit: 273 GB/s is the same as a DGX Spark and roughly a quarter of an RTX 4090. VLA-Perf finds Thor is memory-bound even on the vision encoder, unlike every desktop or datacenter GPU.

What runs on Thor, and how fast

Model	Size	Hardware	Rate	Source and caveat
GR00T N1.7 (Cosmos-Reason2-2B backbone plus SigLIP2 plus flow-matching DiT)	3B	AGX Thor 128 GB	8.9 Hz PyTorch, 12.4 Hz TensorRT full pipeline	NVIDIA Isaac-GR00T hardware recommendation; "10+ Hz recommended minimum"; action chunking sustains about 30 FPS execution
GR00T N1.7	3B	AGX Orin 64 GB	2.9 Hz PyTorch, 6.6 Hz TensorRT	Same; "only suitable for slow, non-reactive tasks"
GR00T N1.7	3B	H100 or RTX Pro 6000	30+ Hz, 35.9 Hz max on H100	Hugging Face model card
GR00T N1.6	3B	AGX Thor	22 to 24 Hz (41 to 45 ms)	An independent developer with hand-written CUDA kernels on the NVIDIA forum; precision and power mode unstated [single source]
pi0.5	about 3B	AGX Thor	23 Hz (44 ms)	Same forum thread [single source]
pi0	about 3B	AGX Thor	22 Hz (46 ms) in the same thread; 19.0 Hz (52.6 ms) in VLA-Perf	Two independent measurements agree at roughly 20 Hz
pi0	about 3B	RTX 4090 or 5090	32 Hz and 54 to 57 Hz	VLA-Perf; forum
pi0-class VLA scaled to 16.7B	16.7B	AGX Thor	2.1 Hz	VLA-Perf; useless for control
VLA with a 100-timestep KV-cache context	about 3B	AGX Thor	8.1 Hz, and 1.3 Hz at 1,000 timesteps	VLA-Perf; long-context memory inside the VLA is not affordable on Thor
Qwen3.6-35B-A3B NVFP4 (MoE, 3B active) via vLLM	35B total	AGX Thor	190 to 239 tok/s aggregate at 16 concurrent; 72 to 84 tok/s single-stream peak; time to first token 30 to 200 s under heavy load	NVIDIA forum benchmark report, Aug 2026 [single source]
Same	35B	DGX Spark	249 to 268 tok/s aggregate; 92 tok/s peak	Same report
Same	35B	RTX Pro 6000 Blackwell	818 to 1,053 tok/s aggregate; 316 to 336 tok/s peak	Same report
Llama 3.1 8B (quantised)	8B	Jetson, device unspecified	44 tok/s single user	Jetson AI Lab benchmarking tutorial example
Nemotron 3.5 Lightning, Qwen3.8-27B	27B class	AGX Thor or Orin	"strong options" per the NVIDIA blog; numbers not extracted	NVIDIA technical blog, page did not render [single source]
Gemini Robotics On-Device 2	undisclosed, Gemma-based	undisclosed; "runs locally without a network"	undisclosed	DeepMind model card. Trusted testers only; struggles with high-DoF robots; "not evaluated on mobile platforms or whole-body control"
SmolVLA	0.45B	Qualcomm IQ-9075 NPU	31 ms/step, about 32 Hz, INT8	AI Review Zones [single source]
Llama 2 13B	13B	Qualcomm IQ-9075	12 tok/s	Qualcomm and DigiKey materials
Llama 3.2 1B	1B	Hailo-10H	30 to 50 tok/s at about 2.5 W	Hailo and community

On Thor a 3B-class VLA delivers 10 to 25 Hz depending on how much kernel work goes into it; an on-robot LLM planner is realistic at 7 to 8B dense (40 to 50 tok/s) or a 3B-active MoE (70 to 80 tok/s single stream), which is adequate for a reasoning step every few seconds but not for a 30 to 70B judge. Anything above about 8B dense on Thor competes with the VLA for the 273 GB/s bus [inference from VLA-Perf and the vLLM report].

VLA-Perf's deployment finding matters for any two-mode design: "server-side inference (even consumer GPU + WiFi) outperforms Jetson Thor on-device; on-device preferred only under extremely poor network conditions (4G or worse)"; a device-server split with KV-cache shipping is slower than either; and asynchronous inference gives 2.6 to 13.8 times speedups over wireless. So a connected mode that runs the VLA on the server is technically faster, but it makes safety depend on the radio, which is why nobody ships it for humanoids.

Alternatives to Thor

Platform	AI compute	Memory	Power	Status for a Singapore robot builder	Notes
Jetson AGX Orin 64 GB	275 TOPS INT8	64 GB, 205 GB/s	15 to 60 W	Safe fallback, available to 2032, USD 2,999 now	A 3B VLA at 6.6 Hz on TensorRT: fine for a wheeled manipulator doing slow pick and place, not for a reactive biped
Qualcomm Dragonwing IQ-9075	100 dense and 200 sparse INT8 TOPS NPU; 1.2 TFLOPS GPU	up to 36 GB LPDDR5	about 30 W fanless envelope	EVK on DigiKey and Lantronix; module price undisclosed; powers Booster and VinMotion humanoids	Strong performance per watt; the QNN and AI Hub toolchain is less mature than CUDA for VLAs; SmolVLA at 32 Hz shows small VLAs fit
Hailo-10H	40 TOPS INT4, 20 INT8	4 to 8 GB LPDDR4X on module	2.5 W	An M.2 module; an Israeli company, no export issue into Singapore	Good for a 1 to 2B VLM sidecar for perception and captioning for episodic memory, not for the VLA
Horizon Robotics (D-Robotics) RDK S100 and S100P	80 and 128 TOPS BPU plus 6 A78AE plus 4 Cortex-R52 MCU	typically 8 to 16 GB [inference]	tens of W	Sold via Waveshare, Yahboom and OpenELAB; 50+ customers including bipeds and humanoids; shown at Embedded World 2026	Single-SoC compute plus control with a real-time MCU is attractive for a biped's low-level controller; Chinese-origin, not US-controlled
Horizon Journey 6P	up to about 560 TOPS [prior knowledge]	automotive	automotive	Automotive-only SKUs, 20+ OEMs	Not a robot part yet
Black Sesame A2000X	about 1,000 TOPS "equivalent" [single source]	not applicable	not applicable	Announced for embodied AI; no retail path seen	Watch list
Rockchip RK3588	6 TOPS NPU	up to 32 GB	5 to 15 W	Ubiquitous and cheap; used in quadrupeds and service robots	A perception or utility board, not a brain
Rockchip RK3688	32 TOPS NPU, 12 Armv9.3 cores, LPDDR6 200 GB/s	LPDDR6	about 10 to 20 W [inference]	New product introduction mid-2026, volume late 2026	An interesting 2027 auxiliary board
Huawei Ascend 310, Atlas 200I A2	16 TOPS; the Atlas 200I A2 is 8 TOPS with 4 GB	small	low	Available in China	Ascend 910B, C and D use "anywhere in the world" was flagged by BIS in May 2025 as an export-control violation; the 310 is not named, but reputational and partner risk remains. Avoid for a company that wants US and EU customers [inference]
Xpeng Turing (3 chips, 2,250 TOPS in IRON)	2,250 TOPS aggregate	not applicable	not applicable	In-house only; IRON mass production targeted for end-2026; a USD 900M raise in Aug 2026	Not purchasable
Intel Core Ultra and AMD Ryzen AI Embedded	40 to 50 TOPS NPU plus iGPU	DDR5	15 to 45 W	Widely available	Not researched this session; fine for a site edge box, not competitive with Thor for a VLA [prior knowledge]

Site-server options for a connected mode

Option	Compute	Memory	Price	Fit
NVIDIA DGX Spark (GB10)	about 1 PFLOP FP4	128 GB unified, 273 GB/s	USD 3,999 at launch in Oct 2025, USD 4,699 since Feb 2026 (Newegg 4,399; Best Buy 5,404)	Runs up to about 200B-parameter inference but is bandwidth-limited: 92 tok/s peak on a 3B-active MoE, only about 1.3 times a Thor. A good dev box and small-site memory and planner server for 2 to 4 robots [inference]
One or two RTX Pro 6000 Blackwell (96 GB GDDR7 each) in a 2U server	about 4 PFLOPS FP4 per GPU	96 GB per GPU, about 1.8 TB/s	GPU roughly USD 8 to 10k each [prior knowledge]; server USD 20 to 35k	316 to 336 tok/s single-stream and 800 to 1,000 tok/s aggregate on Qwen3.6-35B; GR00T fine-tuning with 30+ Hz inference. The right site brain for 5 to 20 robots: a planner LLM of 30 to 70B, a judge, memory consolidation and nightly fine-tunes
A cluster of Jetson Thor dev kits	2 PFLOPS FP4 each	128 GB each	USD 5,499 each	The same memory-bandwidth ceiling as a single robot; only sensible where an identical software stack on server and robot is required. Not recommended as the primary site server [inference]
Cloud (Gemini Robotics ER 2 via the Live API, or hyperscaler GPUs)	not applicable	not applicable	per token	Sub-second per call but internet-dependent; hospitals and ports in Singapore will often forbid it for video; an optional tier only

Network latency budget for robot-to-server calls

Link	Measured latency	Source
Private 5G SA, lab	sub-1 ms downlink median, 1.1 ms at the 99.9th percentile	Adamuz-Hinojosa et al. via Firecell
Private 5G, factory floor, low load	3.9 ms uplink and 4.6 ms downlink one-way	Silicon Austria Labs 2023
Private 5G, campus	11.36 ms average round trip, 33.1 ms max	Toshiba Europe 2024
Private 5G, safety-function testbed	20.4 ms average round trip; 99 ms at the 99th percentile for the e-stop path; 149.6 ms worst case including VPN and PLC	Helmut-Schmidt University 2024 (arXiv 2407.15177)
Private 5G, heavy load	25 ms mean with QoS, 53 ms without	Firecell summary
Private 5G, range across studies	7 to 28 ms	Shah et al.
Wi-Fi 6E against public 5G NSA, AMR video and audio	Wi-Fi 6E video 123 ms, audio 13.9 ms; 5G NSA video 205 ms, audio 33.7 ms; 5G to cloud video 244 ms	Cambium Networks field test [single source]
Rule of thumb	50 ms round trip at 20 km/h is 30 cm travelled before a command lands; 10 ms is 3 cm	Cambium

Wi-Fi 6E roaming measurements under 802.11r, k and v were not retrieved this session; industry practice is 50 to 200 ms handoff gaps unless fast roaming is tuned [prior knowledge]. The working assumption is 10 to 30 ms typical round trip and 100 to 150 ms tails on either private 5G or Wi-Fi 6E, with occasional multi-second dropouts during roaming or lift transits.

The realistic split

Function	Where	Rate / latency	Why
Joint servo loops, whole-body controller for biped balance	On-robot, MCU plus Thor CPU cores, or an RDK S100-class R52 MCU	500 to 1,000 Hz servo; 100 to 500 Hz whole-body control	Hard real time; any radio dependence is unsafe
Safety monitors (e-stop, geofence, speed and separation, collision)	On-robot, an independent safety controller	100 Hz or more, sub-10 ms reaction	The standards (ISO 10218, 13482, 3691-4) assume local safety; the 5G e-stop path shows 99 to 150 ms tails
VLA or manipulation policy, 3B class	On-robot Thor	10 to 25 Hz replanning; 30 to 50 Hz action execution via chunking	Measured on Thor; a server VLA is faster but radio-dependent
Local perception, SLAM, scene graph (working memory)	On-robot	Sensor rate; graph update 1 to 10 Hz	Needed in Solo mode
Local planner or fallback LLM (7 to 8B dense, or 3B-active MoE)	On-robot	40 to 80 tok/s; one reasoning step per 2 to 5 s	Enables Solo mode with reduced task generality
Episodic memory capture (captioning at about 1 Hz, embedding)	On-robot	1 Hz	Cheap; the ReMEmbR pattern on Orin proves it
High-level task planner (30 to 70B), LLM judge or verifier, VLM long-horizon QA	Site server, RTX Pro 6000 class	0.5 to 3 s per call; called only at task boundaries or on exception	Cannot fit next to the VLA on Thor; latency tolerant
Long-term memory consolidation, site KG, provenance, contradiction resolution	Site server	Seconds to minutes, batch	Multi-writer reconciliation needs one authority
Fleet arbitration (traffic, lifts, doors, task DAG allocation)	Site server, Open-RMF core plus a task allocator	100 ms to 1 s	Matches RMF and DeepFleet practice
Skill library, policy versions, nightly fine-tunes, sim-in-the-loop	Site server or cloud	Hours	The AgiBot fleet RL pattern; a GR00T fine-tune needs 40 to 80 GB VRAM per GPU

A local-only mode is therefore everything above the line: the robot keeps working with a smaller planner, a 30 to 60 minute episodic buffer, a read replica of the site knowledge graph and conservative traffic rules. A connected mode adds the big planner, the judge, the fused scene graph and fleet arbitration, and it never takes over the control loop or safety.

Part D. Supply-chain constraints for a Singapore buyer

This covers on-robot and edge-server compute only.

Jetson Thor and Orin classification. NVIDIA states ECCN 5A992.c for Jetson modules and developer kits. That is a mass-market encryption classification, licence-free to Singapore, with no know-your-customer requirement beyond distributor terms. No source was found placing Thor under 3A090 or 4A090. [inference: Thor's 2,070 sparse FP4 TFLOPS is below the datacenter-GPU total processing performance thresholds and NVIDIA markets it globally.]
Chip Security Act (H.R. 3447 and S. 1705). Passed House Foreign Affairs 42-0 on 26 Mar 2026; not enacted as of the sources read. It covers ECCNs 3A090, 3A001.z, 4A090 and 4A003.z (H100, H200, B200, GB200, MI300X and systems containing them), with location verification within 180 days of enactment. It does not mention 5A992 or Jetson. It would touch an RTX Pro 6000 Blackwell site server, a 3A090-class part [inference], and possibly a DGX Spark; Singapore importers are explicitly called out as facing "particular scrutiny".
BIS January 2026 rule. H200 and MI325X to China moved to case-by-case with a 25% tariff, a 50% volume cap, third-party testing and know-your-customer checks. It is not applicable to Singapore end-use, but distributors now apply know-your-customer checks and end-user statements to Blackwell workstation and server GPUs everywhere [inference]. The AI Diffusion Rule that would have put Singapore in Tier 2 was rescinded in May 2025.
Huawei Ascend. BIS guidance of May 2025 says using Ascend 910B, C or D "anywhere in the world" risks violating US export controls, and the guidance also covers "comparable" Chinese chips. Ascend 310-class edge parts are not named, but for a Singapore company selling into hospitals with US-affiliated vendors and into ports with US customers, any Ascend part is a commercial liability.
Chinese alternatives that are clean to import into Singapore: Horizon and D-Robotics RDK S100, sold openly via Waveshare, Yahboom, OpenELAB and ThinkRobotics; Rockchip RK3588 and RK3688; and Black Sesame when a retail path appears. China has no export restriction on these SoCs to Singapore; the risk is future US "comparable chip" language and customer perception, not legality [inference].
Singapore rules. The Strategic Goods (Control) Act applies, with the Strategic Goods (Control) Order 2025 in force from 1 Dec 2025 and the control list last updated 23 Feb 2026. Permits are required for export, transhipment, transit, intangible technology transfer and brokering of listed items. Plain import for domestic use is not a permit activity, but a foreign exporter may require an International Import Certificate from Singapore Customs, and Customs offers Preliminary Advice where it is unclear whether an item is listed. Re-exporting robots containing controlled GPUs, for example a Thor-equipped robot sold to Vietnam or Indonesia, is the case to plan for [inference].
Lead times and pricing. The July 2026 Jetson price list (T5000 USD 4,999, T4000 USD 2,999, Orin 64 GB USD 2,999 at 1KU) can be expected to hold or rise while LPDDR5X lead times sit at 26 to 58 weeks, and volume module orders are non-cancellable and non-returnable. Dev kits have been intermittently out of stock in the US marketplace, and APAC distributors (Seeed, Waveshare, Connect Tech, Leetop, Silicon Highway) are the practical channel.

What is not feasible with current chips

One brain running the VLA for many robots from a server over wireless for a biped. VLA-Perf shows it is faster in the lab, but 5G and Wi-Fi tails of 100 to 150 ms and dropouts make it unsafe for balance and manipulation; no humanoid vendor ships it.
Real-time shared weights, or hive learning where what one robot learns is instantly usable by the others. Fleet updates are minutes to nightly: AgiBot pushes every 50 training steps, and Boston Dynamics' "immediately replicated" is a checkpoint push, not live.
A 30 to 70B reasoning model on the robot alongside the VLA. Thor's 273 GB/s bus gives 2 Hz on a 16.7B VLA and about 80 tok/s on a 3B-active MoE; a 70B dense model is a server job.
Long-context memory where the VLA remembers the whole shift on-device. 100 timesteps of KV cache drops Thor to 8 Hz and 1,000 timesteps to 1.3 Hz. Memory must be external and retrieved, not in-context.
Peer-to-peer consensus memory across dozens of robots with no server, and a cross-site global brain with consistent state. Neither is demonstrated anywhere, and the consistency problems of leakage, stale propagation, contradiction and provenance are only now being formalised.
Gemini Robotics On-Device 2 as a biped's controller. DeepMind states it is not evaluated on mobile platforms or whole-body control and that it struggles with high-DoF robots.
Any claim of running on Huawei Ascend, or a dependence on a single Chinese SoC for the core brain, where US or EU customers are in scope.

What this means for Tidewell

Nothing in this report changes what we have already published. Tidewell Brain has two modes: Solo runs everything a robot needs on the robot, including the manipulation policy, the safety monitors, a local planner, working memory and a rolling episodic buffer, and is a first-class state rather than a fallback; Crew connects every robot on the site network to one governed memory with core, short-term and long-term tiers, promoted into core memory only through a named human. The compute split is the one the measurements above describe: Jetson Thor T5000 on the biped and T4000 on the wheeled manipulator with an Orin cost-down variant, and one RTX Pro 6000 class site server for the task planner, the verifier, memory consolidation and fleet arbitration. What we do not claim is unchanged: no manipulation policy served over wireless, no weights updating across the fleet in real time, no 70-billion-parameter model on the robot, no whole shift held in the policy's context, no cross-site global brain, and no Huawei Ascend anywhere in the stack.