AI-assisted validation for warehouse robot releases
Find the failures that only appear after the software changes.
AMPJ is building a reusable validation platform that uses GPU-accelerated simulation and AI-assisted failure analysis to help warehouse robotics teams reproduce critical failures, compare software builds, and review regressions before deployment.
Initial development focus: ROS 2 warehouse robots validated in NVIDIA Isaac Sim.
Design Partner ProgramProduct visualization · illustrative scenario
10-second product visualization
Same scenario. Different behavior.
A software change can alter how a robot behaves in the same environment. AMPJ is designed to surface that regression before deployment.
Robot releases create behavior that unit tests cannot see.
Changes to perception, navigation, controls, learned policies, sensor configuration, or vehicle dynamics can resolve one issue and introduce another. The most dangerous regressions are often intermittent, cross-system, and discovered too late in physical testing.
The gapA successful demonstration is not the same as a validated release.
AMPJ Validation System
One system from candidate build to engineering review.
A reusable software layer for defining validation campaigns, executing controlled scenario variations, reproducing failures, comparing robot builds, and preparing structured evidence for engineering review.
- Define
- Baseline build, candidate build, operating domain, environment, and criteria.
- Execute
- Controlled scenario variations across occlusion, speed, payload, degradation, and traffic.
- Detect
- Threshold violations, collisions, near misses, deadlocks, mission failures, and behavioral changes.
- Reproduce
- Preserve the exact scenario and determine whether the failure is repeatable.
- Compare
- Show what changed between the baseline and candidate builds.
- Review
- Prepare structured evidence for a human engineering release decision.
AMPJ is developing a reusable software product. Technical pilots are used to validate product requirements and integration workflows; AMPJ is not a consulting or outsourced robotics development firm.
Turn thousands of low-level signals into a small number of engineering questions.
AMPJ combines deterministic safety criteria with AI-assisted analysis. Deterministic checks measure what happened. AI-assisted analysis helps engineers understand which failures are related, which regressions deserve attention, and what should be tested next.
Deterministic layer
- Collision and minimum-separation checks
- Detection and stopping thresholds
- Mission and navigation checks
- Reproducibility evaluation
- Candidate-versus-baseline comparison
AI-assisted layer
- Failure groupingIn development
- Regression prioritizationIn development
- Signal correlationPlanned
- Engineering summariesPlanned
- Scenario-suite recommendationsResearch direction
AI assists analysis and prioritization. It does not certify a robot as safe or make the final release decision.
Evidence engineers can act on.
WPS-1842Illustrative output · synthetic warehouse scenario
Scenario
A pedestrian emerges from behind warehouse racking while the reference AMR operates under a simulated payload.
Behavioral difference
The candidate build detects later than the baseline and begins braking with reduced separation.
Engineering evidence
The failure is reproduced, preserved, compared, and marked for human review.
Illustrative build comparison · scenario WPS-1842
| Measure | Build 0.8.13 | Candidate 0.8.14 | Change |
|---|---|---|---|
| Detection latency (WPS-1842) | 512 ms | 684 ms | +172 ms |
| Minimum separation (WPS-1842) | 0.74 m | 0.31 m | −0.43 m |
| Review status | Baseline | Review required | — |
Synthetic values shown for illustration only. Real comparisons are produced against customer-defined criteria.
Illustrative failure record from a synthetic scenario. Nothing on this page reads from or writes to a backend.
Designed for NVIDIA-accelerated Physical AI workflows.
A layered validation architecture: robot software in the loop, GPU-accelerated simulation underneath, and AMPJ failure intelligence between simulation output and the humans who make the release decision.
- ROS 2
- Navigation
- Perception
- Controls
- NVIDIA Isaac Sim In development
- OpenUSD environments In development
- Synthetic camera & LiDAR generation
- NVIDIA RTX Planned architecture
- CUDA Planned architecture
- Scenario management
- Deterministic checks
- Failure grouping
- Build comparison
- Evidence generation
- Human review
- Release decision
- Isaac Lab — scenario randomization and learned-policy evaluation Research direction
- NVIDIA Cosmos — physical AI scenario generation and reasoning Research direction
Technology names describe AMPJ’s technical development direction. They do not imply NVIDIA endorsement, partnership, program membership, or acceptance into NVIDIA Inception.
Building the first warehouse release-validation workflow.
Active development, in the open about what exists and what is next. If real prototype media is configured it appears here; otherwise this build ledger reflects current work.
A bounded warehouse scene for repeatable validation scenarios.
A reference edge case for detection and stopping evaluation.
Threshold-based evaluation of detection, stopping, and separation.
Candidate-versus-baseline behavioral comparison.
Organization of related failures across controlled runs.
Validate the release question your current workflow cannot answer.
Tell us about the robot software build, simulation environment, and bounded failure mode your team needs to evaluate.
Direct email · info@ampj.net
A technical pilot starts from one bounded validation objective, a baseline and candidate build, and an existing or reference simulation environment. Pilot scope, commercial terms, and schedule are defined after a technical fit review.
Data requirements, retention, and deployment constraints are agreed during pilot scoping — only inputs needed for the scoped objective are requested, and customer data is never sold or used for advertising.
Design Partner ProgramCommon questions from robotics teams.
No. AMPJ is a validation and release-review layer designed to work with simulation environments. It organizes scenario execution, failure reproduction, build comparison, and engineering evidence around a specific release question.
Warehouse autonomous mobile robots, autonomous forklifts, and related ROS 2-based mobile robot systems. Other robot categories may be evaluated during technical fit.
Deterministic criteria identify threshold violations and measurable regressions. AI-assisted analysis helps group related failures, prioritize review, correlate engineering signals, and summarize important behavioral changes.
No. AMPJ provides simulation-based validation evidence to support engineering review. It is not a regulatory or safety certification authority, and final release decisions remain with the customer.
Typical inputs include a robot model or reference configuration, a ROS 2 stack or simulation interface, baseline and candidate builds, relevant operating constraints, and one bounded validation objective. Exact requirements are agreed during technical fit.