M2MTD SYNNEX USA warehouse Physical AI POC
From perception to physical action, at the edge.
A practical path from constrained warehouse POC to repeatable Physical AI architecture.
Start with one proposed workcell and a measurable outcome.M2M helps bring Physical AI into warehouse operations.
MEA connects edge perception with existing equipment and enterprise workflows, while preserving controller ownership.
MEA GX-100 · edge node
Reference configurationThe MEA GX-100 is the on-site node that runs perception next to the equipment and carries the physical interfaces used to connect to it. Figures above are the reference configuration, not a benchmark or a promised throughput.
01The operational problem
Typical brownfield systems
Availability and relevance at the selected site are confirmed during discovery.
Conveyors and sorters
Fixed and mobile robots
WCS and WMS
ERP and enterprise workflow
PLCs and control cabinets
Photoeyes, scanners and sensors
Legacy OT with long service life
Manual exception handling
Useful perception and reliable integration have to work together: from what a camera sees to what a controller or a person is permitted to do about it.
MEA proposes an edge layer around the stack already running the building, with interfaces and operating boundaries agreed during discovery.
Proposed application and outcome
Start with randomized bin picking, pending team and TD SYNNEX selection. Measure verified pick outcomes, retries, successful picks per hour and investigation time against an agreed baseline.
Core story
02Proposed first workcell
Proposed for the first evaluation, subject to team and TD SYNNEX selection: representative cartons, polybags, cylinders, irregular shapes, overlaps, reflective film and transparent packaging in a bounded workcell.
Illustrative perception output
No robot connected
All classes, depths, confidence scores and actions below are interface examples, not measured results or robot commands.
Sensor
RGB-D, fixed mount, tote framed
Low confidence
2 objects held for re-perception
Action
Illustrative preferred grasp 0.91
Proposed task: evaluate whether instance masks can separate overlapping representative products, including difficult polybags and reflective film.
Confidence handling
0.91 illustrative confidence / gate 0.65
Scope discipline
Model choices, gripper constraints, controller interfaces and the acceptable operating envelope are confirmed and measured against representative products before any production decision.
03Verification and recovery
Reliability comes from verification, recovery and graceful exception handling.
01
Attempt
Robot executes the ranked grasp
02
Fail
Vacuum loss, slip or empty transfer
03
Detect
Verification catches the failure in cycle
04
Re-perceive
Scene re-captured after tote disturbance
05
Alternate grasp
Next ranked candidate, different approach
06
Retry
Bounded retry budget per item
07
Exception
Human handling when the budget is spent
04Models, data and validation
Model names and versions remain selection inputs. Initial evaluation is separate from any later, approved improvement round.
Proposed stage 01
Low confidence picks, repeated retries and human exceptions would be logged with the frames that produced them, so the gap is measured rather than guessed.
What it would produce
Flagged cycles per shift
Initial evaluation selects and measures the starting approach. Later improvement rounds use separately approved capture, labelling, training, validation and rollout with site agreement.
Reference approach
NVIDIA’s published FOX blueprint describes a comparable model lifecycle built on TAO, and informs the approach above. Scope is agreed per POC, and the training and validation environment is sized separately from the edge node.
View NVIDIA source(opens in a new tab)| Stage | Model / version | Approach | Inputs | Outputs | Developer | Model owner | Evidence |
|---|---|---|---|---|---|---|---|
| Detection | Selection pending team confirmation | Pretrained starting point; custom work only if evidence shows a gap | RGB images | Locations, classes, confidence | To agree: M2M, partner or OEM | To agree | Proposed POC validation |
| Segmentation | Selection pending team confirmation | Pretrained or fine-tuned instance segmentation | Representative tote images | Per-item masks | To agree: M2M, partner or OEM | To agree | Proposed POC validation |
| Depth / pose | Selection pending team confirmation | Method matched to the selected camera and items | Depth and aligned RGB where available | Surface geometry and candidate pose | To agree: M2M, partner or OEM | To agree | Proposed POC validation |
| Grasp ranking | Selection pending team confirmation | Ranking against gripper and workcell constraints | Masks, pose, gripper constraints | Ranked candidates and confidence | To agree: M2M, partner or OEM | To agree | Proposed POC validation |
| Verification | Selection pending team confirmation | Post-pick visual and available station-signal checks | Post-action image and station signals | Verified transfer, retry or exception | To agree: M2M, partner or OEM | To agree | Proposed POC validation |
| Step | Proposed evidence |
|---|---|
| Initial evaluation | Representative SKUs, materials, lighting, occlusion and difficult packaging; collection and site retention agreed first. |
| Data quality | Label definitions, sampling, label QA and traceability agreed before scoring. |
| Model work | Training or fine-tuning only where the selected starting point does not meet the acceptance envelope. |
| Held-out test | Unseen SKUs and difficult packaging excluded from training and used for final evaluation. |
| Operational measures | Failure classes, first-attempt success, retry outcomes and human exceptions compared with baseline. |
| Acceptance | Target, evaluation period and acceptance owner confirmed before testing. |
| Later rounds | New failure examples can inform a separately approved improvement round after initial evaluation. |
| Measure | Expected target | Measured result | Conditions to record |
|---|---|---|---|
| End-to-end latency | Target to agree | Evaluation planned | Camera, resolution, model version, precision, power, thermal state and runtime |
| Perception throughput | Target to agree | Evaluation planned | Frames or scenes per second; not robot picks per hour |
| Successful robot picks / hour | Baseline and target to agree | Evaluation planned | Product mix, retries, motion, gripper and exception time |
| Memory use | Operating envelope to agree | Evaluation planned | Models, camera concurrency, resolution, precision and retained buffers |
TOPS is a compute rating, not application performance.
| Capability | Evidence shown | Next status |
|---|---|---|
| Warehouse perception | Interactive illustration | Proposed POC validation |
| Robotics | Interactive illustration; OEM controller ownership preserved | Proposed POC validation |
| Training / improvement | FOX-informed design reference | Later approved round |
| Simulation | Interactive scenario illustration, not validated robot physics | Applicability assessment |
| Video intelligence | Recorded other-environment example | Proposed warehouse evaluation |
| Production sensing / edge integration | Prior production evidence in another environment | Reference for delivery experience |
Industry references
Scroll the diagram sideways to read the labels
Industry reference · conceptual illustration by M2M
Generated training data
A published example of generated and annotated scene variants used to build a training set for an industrial vision model.
Comparable to the proposed step where rare packaging and lighting conditions are recreated instead of waited for.
View NVIDIA source(opens the official article in a new tab)Data retention, labelling and any training round are agreed with the site. FOX informs the lifecycle design only; it does not establish MEA GX-100 compatibility or performance.
05Decision, action and ownership
A warehouse running several Physical AI applications needs them reconciled at the edge, under site rules, with a clear handoff to a person when confidence runs out.
Site coordination layer
/Reads live state from perception, robots and the control layer
/Applies site rules and priority before any action is proposed
/Keeps a written record of what was decided and why
/Hands anything outside the envelope to a named human
Coordination scope for the POC is limited to the agreed workcell and observed zones.
Pick agent
Randomized bin picking in the bounded workcell.
Material transport agent
Tote and pallet movement between staging, pick and dock.
Verification agent
Did the physical action actually happen as instructed.
Safety awareness agent
Person and forklift proximity in shared travel space.
Asset health agent
Conveyor drives, motors and gearboxes on the observed line.
Extension
Other applications may join the same coordination pattern after the first workcell establishes interfaces, evidence and operating ownership for the site.
Proposed first-workcell example
Illustrative and proposed · nothing here triggers a real action
Detection
Repeated failed cycles on the same bin, logged with the frames that produced them.
Evidence
Attempt records, grasp scores and the post-attempt frames are kept together.
Investigation
The coordinator checks whether the failures share a packaging class or a bin position.
Proposed action
Proposes re-perception with an alternate grasp, or a human exception if the class is unknown.
Human approval
The cell owner approves the retry policy change or takes the exception.
Handoff
The approved request goes to the robot controller, which keeps motion and safety ownership.
Verification
The next cycles are measured against the same failure class and added to a review backlog.
MEA role in this example
The edge node would run detection and grasp ranking in the proposed cell and record each attempt and its result.
Open for POC validation
Cycle-time budget, controller interface, and how many retries are allowed before a person is called.
Motion and safety functions stay with the existing robot and safety controllers.
Industry references
Scroll the diagram sideways to read the labels
Industry reference · conceptual illustration by M2M
Material transport
A published example of coordinated material movement between stations on a factory floor.
Comparable to the transport agent scope proposed here, at a different site and scale.
View NVIDIA source(opens the official article in a new tab)
Scroll the diagram sideways to read the labels
Industry reference · conceptual illustration by M2M
Procedure verification
A published example of checking that a documented work procedure was actually followed.
Comparable to the verification agent scope: confirm the physical action happened as instructed.
View NVIDIA source(opens the official article in a new tab)06Application architecture and hardware
Proposed warehouse AI architecture
Architecture informed by NVIDIA’s FOX blueprint
Scroll the diagram sideways to read the labels
Layer 01 · Warehouse equipment
Cameras, sensors, controllers
Existing site equipment
Layer 02 · MEA GX-100 edge layer
MEA GX-100
Reference configuration
Layer 03 · Operations orchestration
Operations orchestration
Separately sized central compute
Stays with the existing owner
Robot controllers and safety controllers keep motion execution and safety functions. The edge layer requests and records; it does not perform safety-rated stopping.
Scope and POC validation
MEA GX-100 provides the edge layer. Central AI services are sized separately. Integration and performance would be validated against the agreed POC scope.
View NVIDIA source(opens in a new tab)Core path
Cameras / sensors
RGB, depth, vibration, temperature, current, existing OT signals
MEA GX-100 edge perception
On-site detection, segmentation, depth / pose, grasp ranking and verification for the proposed picking workcell.
Video investigation layer
Proposed warehouse evaluation
Video intelligence
NVIDIA Metropolis VSS is proposed as a separately sized investigation layer that receives edge video and event references, then returns summaries and clip references for supervisor review. Controller and safety ownership remain unchanged.
Full video intelligence detail and the recorded demonstrationPolicy, confidence and human approval
Site rules, confidence gating and an operator decision point before any action is requested
Existing WMS / WCS / PLC / robot controllers
Action requests into systems that already run the floor. Motion and safety stay with the OEM controller.
Verification, evidence and recovery
Outcome checks, retained event records, retry and escalation
Proposed first workcell: randomized bin picking, pending team and TD SYNNEX selection. Edge perception can run on the MEA GX-100; the complete video intelligence stack is sized separately.
MEA GX-100 interfaces, compute and storage detailOwned by
M2M
Owned by
TD SYNNEX
Owned by
Existing OT
Owned by
Robot OEM
Interfaces are not integration
The MEA GX-100 provides the physical connection paths. Protocol and controller compatibility is confirmed during discovery, and the robot and safety controllers keep motion and safety ownership.
MEA hardware
The proposed edge-perception workload can be evaluated on the MEA GX-100. The complete VSS, VLM and search infrastructure must be sized separately based on camera count, models, latency, retention and workload requirements.
Enclosure, cable routing, power conditioning, site airflow and the final mounting position are selected for the work area. Cellular use may require thermal derating.
Compute module
NVIDIA Jetson AGX Orin 64GB
AI performance
Up to 275 TOPS (reference figure, not a throughput guarantee)
Memory
64GB 256-bit LPDDR5 DRAM
Networking
4 x 10/100/1000 Mbps; PoE optional (IEEE 802.3af/at)
USB
4 x USB 3.2 Gen 2 (10 Gbit/s), 2 x USB 2.0
Industrial I/O
4-ch DI, 4-ch DO; serial ports optional
Onboard storage
64GB eMMC; NVMe M.2 2280 optional, not installed at base
Power input
9 ~ 36 VDC, 16-4A
Cooling and range
Fanless; -10 ~ +60 °C with 0.7 m/s airflow (MaxN mode)
Reference platform specification for the MEA GX-100. It is not a confirmed installed bill of materials; the supplied configuration is agreed per deployment.
Reference specification
0 of 7 specification groups expanded
NVMe storage is an option selected per deployment.
SIM slots do not mean a cellular modem is installed.
GMSL channel count is confirmed with the selected configuration.
Range notation reproduced exactly as printed in the source.
Depends on the installation situation and interface connection. Derating of the maximum operating temperature is possible when a cellular LTE module is used.
Dimensional drawing figures, kept separate from the specification table figures.
The DIN rail mounting kit is optional and is not included.
Certifications do not apply to all configurations. Applicability is confirmed for the selected configuration.
Source document last updated 21-Apr-2026. All specifications are subject to change without notice.
Evidence storage and connectivity
Sensing layer
Dual track edge processing
Track A
High frequency stream
Mechanical vibration, sampled densely enough to catch micro anomalies, pre processed on the device so no raw stream has to leave the floor.
Track B
Low frequency stream
Temperature and RPM on a slower cycle, with sampling selected and validated for the measured phenomena and available processing headroom.
The monitoring path is designed to run alongside production equipment rather than inside its control loop, so intelligence is extracted without disrupting the process. The exact integration is agreed per asset with the equipment owner.
Protocol / vendor support to be confirmed during POC discovery. No specific controller, network or system compatibility is claimed ahead of site assessment.
07Operator investigation
Ask MEA answers from structured and text event records within a displayed source window. The VSS reviews video, returns timestamps and supporting clips, and keeps the shift supervisor responsible for disposition.
Evidence center
Loading feed
Reading the event store Latest reading (event time) unknown.
Reading the event store...
Ask MEA, event-log assistant demonstration
Reading the event store
Loading feed
Ask the warehouse
what happened.
A demonstration of plain language questions answered from stored event records. Answers are generated from the event text by a hosted language model, not from video search and not from an on device model. Record references are shown with every answer; playable footage is not part of this demonstration.
08POC acceptance
01
Read only deployment. Measure detection accuracy, latency and edge reliability under real warehouse conditions.
02
Perception feeds a decision layer that proposes an action. Execution stays supervised and is evaluated against an agreed operator baseline.
03
Closed loop physical action in a bounded workcell, verified every cycle, with defined recovery and exception paths.
Proposed acceptance envelope
Pick success
Verified first-attempt and post-retry outcomes on the held-out SKU sample.
Throughput
Successful robot picks per hour, kept distinct from perception throughput.
Recovery
Failure classes, retry results and human exceptions within the agreed budget.
Investigation
Supervisor time and evidence usefulness compared with an agreed manual baseline.
What the POC needs to start
Site / environment
Representative area, power, network and mounting positions.
Constrained use case
One workcell or flow path with a clear success definition.
Representative SKU sample
Real packaging variety, not best case items.
Cameras
Existing coverage where usable, plus placement for the target zone.
Robot if applicable
Cell, controller access and OEM engagement for the acting phase.
MEA hardware
Edge compute and supporting infrastructure supplied by M2M.
WCS / PLC interfaces
Interface points and signal definitions, confirmed during discovery.
Data access
Reference data to evaluate accuracy and business outcome.
Integration requirements
Security, network and change control expectations.
Measures, baselines, targets, evaluation period and acceptance owner are agreed before evaluation. No improvement figure is claimed in advance.
09Warehouse walkthrough
Randomized bin picking is highlighted as the proposed first workcell, pending team and TD SYNNEX selection. Explore the other zones as additional applications, not committed scope.
Illustrative TD SYNNEX Brownfield Distribution Center. Conceptual facility model for POC discussion. Not an actual TD SYNNEX facility layout.
Each zone panel maps an MEA GX-100 role to the proposed application. The canonical hardware specification remains in the architecture section.
10One edge platform, multiple use cases
01
first attempt pick success and cycle time.
02
material consumed per carton and damage related rework.
03
unplanned downtime minutes.
04
near miss frequency and response time.
05
false reject rate and inspection coverage.
06
unplanned line stop minutes per shift.
11Predictive maintenance
Select a drive
MTR-114
Health score
62
Vibration
4.8 mm/s RMS
Bearing temperature
61 C
Current draw
12.4 A
Illustrative condition trend
Bearing signature drifting over the last 9 days
Illustrative inspection window
Illustrative: inspect within 14 days
Illustrative abnormal-condition progression
Inspection timing requires validation
Normal operation
RPM, temperature and vibration held inside the learned envelope for this drive.
Mechanical stress
A validated vibration-monitoring method may identify a change from the established operating baseline.
Current phase for this drive
Thermal escalation
A correlated temperature rise can add evidence for review; response timing depends on the validated method.
Seizure
RPM collapses as the bearing fails. This is the outcome the alerting exists to avoid.
Combined signals can support detection of abnormal conditions; warning time depends on the asset and validated monitoring method.
All values and inspection windows are illustrative, not a validated forecast or measurements from a TD SYNNEX asset. The signal pattern is informed by prior M2M edge monitoring work on rotating equipment.
12Worker / forklift safety
Illustrative detections and confidence scores; no live safety system connected.
Event response
Alert
Idle
Evidence
No event captured
Workflow
Awaiting event
Alerting and evidence only. Safety rated machine stopping is not claimed. Safety rated control integration is future scope and integration dependent.
13Simulation / digital twin
Camera placement, robot location, bin randomization, conveyor speed and traffic scenarios can be explored against a scene model before anyone mounts hardware on a live flow path.
Explore → Evaluate → Decide
Twin tooling applicability is assessed per site rather than assumed.
Conceptual simulated twin
Conceptual physical scene
52% simulated twin / 48% conceptual physical scene. Both images are conceptual renders, not an actual customer facility.
Scene
Racking, conveyor, AMR lanes, dock faces
Synthetic variation
Pose, lighting, packaging class, occlusion
Transfer target
Same perception build, edge deployed
Conceptual renders for POC discussion.
Loading feed
Reading the event store
The floor plan above is driven by the event store, not by a script. In live mode the newest reading lights its zone as it lands; scrub or replay to rewind the same readings and argue over a shift after the fact.
Events
0/0
Alerts
0
Devices
0
Reading clock
--
Feed refreshed
--
Loading recorded events…
Replay of loading feed. Not measurements from a TD SYNNEX facility.
14Technology enablement
This is the M2M technology stack. NVIDIA technologies are referenced as enablers where they are technically relevant to the architecture.
Component selection is confirmed per deployment.
15Proven in the field
These photographs document prior M2M production sensing and edge-integration work: edge compute, sensing modules and conditioned power deployed around existing equipment.
Relevant delivery experience. A new environment to evaluate.
The prior deployment demonstrates production sensing, mounting, power and edge-integration experience. It does not establish warehouse vision, robotics or TD SYNNEX application results.
Prior deployment in a different industry and operating environment. Results and conditions there are not a forecast of TD SYNNEX outcomes, and no figure is carried across.
Prior M2M production sensing and edge-integration evidence from another environment. Select any photo for the full view and scope.
16Scale
01
One proposed workcell
A selected work area measured against the acceptance envelope agreed before evaluation.
Expansion is a decision made after the result is reviewed and accepted.
02
Multiple zones, same site
Additional zones join the same edge and coordination layer rather than starting again.
Each new zone revalidates camera and sensor configuration, interfaces, representative data and site constraints.
03
Multiple sites
Accepted interfaces and evidence can become inputs to planning for the next building.
Every site revalidates configuration, interfaces, data and local constraints. Rollout outcomes are not guaranteed in advance.
What can carry forward
An accepted result can inform a reusable reference architecture for hardware, edge software, integration and evidence. Where simulation applies, candidate layouts and workflows can be explored before physical evaluation. Any reseller or customer offer remains a separate decision after the internal result is reviewed.
forSelect one work area, agree on the outcome, and define the first Observe phase.