97.5%
Clean runs in a 1,000-run simulator test.
Autonomous AI Validation
Goodness Technology builds safety validation for autonomous AI systems.
Our first demo shows an adaptive robot controller that plans, recovers, and explains safer behavior in simulation before real-world pilots.
96.5%
Fewer crash-runs than the basic robot in the best tested mode.
975/1,000
Runs finished without collision by the Goodness robot.
Main Idea
Before autonomous AI acts, Goodness tests it.
The robot demo is the first proof point for a broader validation layer: testing, scoring, and explaining autonomous behavior before it reaches real environments.
Plan before moving
The robot samples possible paths, scores risk, and chooses safer motion in real time.
Adapt to the scene
Open space, traffic crossings, and narrow corridors produce different driving behavior.
Remember recent danger
The system treats problem areas with caution instead of repeating avoidable mistakes.
Validate before pilots
Benchmarks and telemetry make behavior reviewable before physical deployment.
Why This Matters
The next AI safety question is action.
As AI moves from chat into agents, robots, vehicles, workflows, and physical systems, the question is no longer just “can the AI answer?”
The question is: should this AI be allowed to act?
Goodness is building the validation layer that tests, scores, and explains autonomous behavior before it reaches the real world.
Benchmark Proof
Same test. Very different result.
In 1,000 randomized simulator trials, the Goodness robot finished clean far more often than the basic controller. Detailed controller internals remain confidential.
| What happened | Basic robot | Goodness robot | Plain-English takeaway |
|---|---|---|---|
| Clean runs | 287 / 1,000 | 975 / 1,000 | Ours finishes clean far more often. |
| Runs with a collision | 713 | 25 | Same test. About 28x fewer bad runs. |
| Total collisions | 1,720 | 28 | The basic robot keeps making the same mistake. |
| Safety interruptions | 561,867 | 19,058 | Ours needs fewer emergency corrections. |
| Can explain itself | No | Yes | Live telemetry shows risk, mode, skill, and recovery state. |
How it was tested
1,000 simulator runs across obstacle gauntlet, crossing traffic, and tight corridor scenarios.
Why it matters
Best tested mode cut collision-runs by 96.5% versus the basic robot.
The honest line
This is strong simulator proof. It is not yet a real-world safety certification.
ULTRA Robo · Ground Autonomy Validation
Ekumen Andino · WebotsPerfect replay.
Harder audit.
Profile 4 guides a differential-drive Andino through moving solid obstacles while visible, permeable sensing fields show proximity—not invisible walls.
The recorded demonstration completed 100 of 100 observed runs. We then widened the test envelope with 100 unique simulation seeds, varying obstacle timing and speed plus small, bounded changes to the starting pose.
ANDINO PROFILE 4 · WEBOTS
Every seeded run reached target hold.
Goal reached with zero solid overlap.
Timeouts and Webots application crashes.
Unique seeds with verified trace hashes.
Repeatability ≠ robustness
Both results are true. They answer different questions.
Yes. All 100 manually observed runs reached the target in the recorded operating condition—a 100% observed success rate.
Mostly. All 100 simulated runs reached the target; 96 completed with no solid-footprint overlap—a 96% audited success rate.
Observed evidence · 100 consecutive runs
Watch the complete 100-run record.
This reel documents the manually observed operating condition in which all 100 runs reached the target. It is shown separately from the unobserved varied-condition simulation audit.
Technical snapshot
A real robot model. A measurable test contract.
Public integration details for technical review—without exposing proprietary controller tuning.
Differential-drive Andino
- Wheel radius
- 33.1 mm
- Wheel separation
- 137 mm
- Maximum wheel speed
- 14 rad/s
- Acceleration limit
- 8 rad/s²
Low-resource audit mode
- Motion state
- GPS · IMU · gyro
- Obstacle proximity
- Webots radar
- Contact authority
- Oriented footprint validator
- Visible blue field
- Permeable · non-physical
Profile 4 navigation
- Global planning
- Fast Marching Method
- Local response
- Bounded motion refinement
- Crossing behavior
- Commit · wait · go
- Actuator authority
- Validated commands only
Seeded Webots validation
- Simulation step
- 64 ms
- Success threshold
- ≤ 0.11 m goal error
- Mission timeout
- 90 seconds
- Evidence
- Per-run trace + SHA-256
The honest result
Completion was perfect. Contact avoidance was not.
Four edge-case runs reached the goal but briefly overlapped a solid box. They remain failures under the zero-overlap rule. The measured collision-free rate is 96%, with a 90.2–98.4% Wilson 95% confidence interval.
Evidence boundary: controlled Windows Webots simulation, not physical-hardware certification. Controller internals remain confidential.
Technology
What changes inside the robot behavior.
The system combines lookahead planning, bounded adaptation, recent memory, and readable state reporting so the robot is not a black box during a demo or review.
It looks ahead
It checks multiple possible moves before choosing where to steer.
It changes style
A tight corridor, crossing obstacle, and open path do not get the same behavior.
It remembers trouble
If a spot caused problems before, the robot treats that area with more caution.
It stays inside guardrails
The learning layer can tune within limits, but it cannot rewrite the controller on a whim.
ULTRA Robo · Autonomous Flight Research
Controlled simulationAutonomous decisions
before motion.
ULTRA Robo demonstrates sensor-driven navigation, predictive safety, command validation, memory, and precision landing in a closed-loop Webots drone simulation.
Most autonomous systems generate a command and immediately send it toward the robot. ULTRA Robo inserts an explicit decision layer between intention and execution, evaluating every proposed movement against live sensor conditions, predicted stopping distance, remembered failure patterns, and mission priorities before it reaches the simulated flight controller.
Sensors→
Waypoint planning→
Predictive safety→
Memory→
Command validation→
Flight control
AUTONOMOUS MISSION · WEBOTS
Watch the autonomous mission
No hard-coded route.
No supplied target coordinates.
A simulated DJI Mavic 2 Pro discovers its destination through camera recognition, generates local waypoints from live sensor observations, navigates a constrained obstacle course, and performs a staged precision landing.
The green marker defines the environment’s destination, but its coordinates are never supplied to the autonomy controller. The drone must detect the elevated beacon through its camera, estimate its location, and continuously plan a safe route toward it.
Watch directly on YouTubeWhat you are seeing
One mission. Eleven observable capabilities.
| Capability | Demonstrated behaviour |
|---|---|
| Visual goal acquisition | The camera detects and estimates the green mission target. |
| Sensor-generated waypoints | Local navigation points are created from live obstacle clearances. |
| Core-goal persistence | Temporary waypoints can redirect the drone but cannot replace its mission goal. |
| Multi-sensor awareness | Forward lidar, 360° lidar, radar, near-field lasers, GPS, camera, and IMU contribute to decisions. |
| Predictive safety | Stopping distance, control latency, sensor age, and closing speed are evaluated before motion. |
| Command validation | Commands may be allowed, modified, or refused before actuation. |
| Failure memory | Hazardous sensor-command combinations can influence later decisions. |
| Momentum-aware recovery | The drone brakes, settles, and turns before resuming forward travel. |
| Precision landing | Descent transitions through fast descent, approach, flare, and stable touchdown. |
| Command deadman | Expired commands produce hover rather than continued motion. |
| ROS 2 integration | Sensor, mission, safety, and validated-command interfaces are separated into standard robotics components. |
A safety-aware command pipeline
Wanting to move is not permission to move.
The actuator interface accepts only validated commands. Proposed commands remain observable for analysis, but they have no direct route to the motors. For a flying robot, an emergency stop means a controlled hover—not cutting propulsion.
- 01 · SenseCamera, lidar, radar, GPS & IMU
- 02 · PlanSensor-generated waypoint
- 03 · ProposeMovement command
- 04 · EvaluatePredictive safety & failure memory
- 05 · ValidateAllow · Modify · Emergency hold
- 06 · ActSimulated Mavic stabilizer
The sensor shield
Proximity awareness, made visible.
The shield displays what lidar and radar perceive. It does not create physical barriers or steer the drone; the autonomy stack independently determines the response.
- Clear
- No object in the nearby safety region.
- Blue
- An object has entered the caution region.
- Red
- An object has penetrated the inner boundary.
Simulation study
300 autonomous trials.
Five controlled profiles.
The fixed-course Webots study included memory and predictive-safety ablations plus matched sensor-stress conditions. These are course-specific simulation results—not claims of physical-flight reliability—and demonstrate repeatability only within the defined world, dynamics, sensors, and fault model.
Webots + ROS 2 integration
A platform-neutral validation core connected to a replaceable robotics stack.
Webots supplies the simulated drone, environment, physics, and sensor data. ROS 2 carries observable messages between sensing, planning, ULTRA Robo’s validation engines, and the flight controller.
Robotics Interfaces
Standardized sensor, mission, proposed-command, validated-command, and safety-state contracts.
Autonomy Planner
Interprets the mission, identifies goals, generates waypoints, and proposes movement commands.
Validation Command Engine
Evaluates each proposed command and may allow, modify, hold, or refuse it before actuation.
Memory Engine
Records relevant outcomes and failure patterns so previous hazards can inform later validation decisions.
Webots Flight Adapter
Translates simulated sensor data and sends only validated commands to the simulated flight controller.
Why it matters
Intelligent action is not enough.
Robots operating around students, patients, older adults, and the public must determine whether their actions remain coherent and safe under current conditions.
- Mission goals remain persistent.
- Local plans adapt in real time.
- Prior failures influence future decisions.
- Stale or unsafe commands fail toward a controlled state.
- Every motor-facing command crosses an observable validation boundary.
Current research status
Completed
- Closed-loop Webots autonomy demonstration
- Camera-acquired mission goal
- Sensor-generated waypoint navigation
- Predictive stopping, validation, and failure memory
- Controlled landing sequence
- 300-trial simulation study
- ROS 2 package, interfaces, and runtime parity testing
Next milestones
- ROS bag recording and deterministic replay
- PX4 software-in-the-loop
- Hardware-in-the-loop
- Propeller-off sensor and latency testing
- Supervised tethered flight
Research prototype. Validated in controlled simulation. Physical hardware validation has not yet been completed.
The contribution is not simply a drone completing an obstacle course.
Technical Snapshot
Simulation stack details for technical review.
These are the current demo and validation settings from the repository. They describe the simulated autonomy stack, not final production hardware.
Planning and Control
- Planning fan
- 31 candidate rays
- Planning cone
- 120 degrees
- Ray range
- 340 simulation units
- Base steering rate
- 7.5 rad/s
Speed and Safety
- Cruise speed
- 120 units/s baseline profile
- Caution speed
- 84 units/s
- Clarify speed
- 36 units/s
- Emergency distance
- 16 units, 0.38s confirm window
Memory and Learning
- Spatial memory
- 240 max entries
- Memory half-life
- 45 seconds
- Experience memory
- 180 reward/risk entries
- Learning boundary
- bounded skill/profile adjustments
Validation Harness
- Benchmark size
- 1,000 randomized runs
- Trial window
- 45 seconds per run
- Simulation step
- 60 Hz
- Scenario set
- gauntlet, crossing, corridor
The robot body in the simulator uses a 13-unit collision radius. Skill profiles include baseline, projectile dodge, corridor precision, and flow runner modes.
Live Demonstration
Watch it run. Then try to break it.
The embedded simulator shows the robot moving through obstacles in real time. Open the full demo for controls, telemetry, scenario changes, and side-by-side behavior.
Phone viewing is supported for quick preview, but the full simulator is best on desktop or tablet.
Start with the baseline
Watch how often a simple controller gets trapped by traffic, corners, or repeated hazards.
Switch scenarios
Use gauntlet, crossing, and corridor cases to see how the robot changes behavior.
Read the telemetry
Look for mode, risk, recovery state, and skill behavior while the robot moves.
Embedded demo could not initialize in this browser context. Open the full demo directly.
Current Stage
Working simulation now. Physical pilots next.
We are being clear about what is proven today, what comes next, and who we want to meet.
Current stage
Working simulation demo and benchmark.
Next stage
Controlled physical robot pilot.
Not claiming yet
Production robot safety certification.
Looking for
Pilot partners, advisors, and early investor conversations.
Who We Want To Meet
People building, funding, or validating autonomous systems.
Robotics founders
Mobile robots, warehouse robots, service robots, and other embodied AI systems.
AI safety researchers
Researchers interested in autonomous system validation, scoring, and explainability.
Investors
Investors focused on physical AI, robotics, infrastructure, and safety.
Enterprise teams
Innovation teams testing AI agents, autonomous workflows, or operational AI risk.
Technical advisors
Advisors in robotics, simulation, controls, validation, and embodied AI.
Market Context
The evidence supports a focused validation wedge—not a giant TAM claim.
Professional mobile robots are an established and growing category. [1] At the same time, research and regulation place increasing weight on system-level testing, uncertainty, obstacle detection, and safe control. [3] [4] [8] Goodness is positioned at the decision boundary: after autonomy proposes an action, but before that action reaches the machine.
Why now
Capability is advancing. Assurance remains hard.
Peer-reviewed research
Foundation models now reach perception, planning, and control.
A 2025 IJRR survey also identifies safety evaluation, uncertainty quantification, real-time performance, and reproducibility as unresolved deployment challenges. [3]
Regulatory trigger · 20 Jan 2027
The EU Machinery Regulation adds explicit autonomous-mobile requirements.
Depending on risk assessment, Annex III requires autonomous mobile machinery to operate in protected zones and/or detect nearby people, animals, and obstacles. Its missing-validation-signal stop rule applies specifically to remote-controlled machinery. [4]
Validation practice
System-level evidence requires more than one test method.
A systematic review screened 10,534 papers and coded 192 studies spanning physical testing, model-based testing, runtime monitoring, formal verification, and model checking. [8]
Company-reported deployment scale
Mobile robots already operate inside large logistics networks.
Amazon reported one million robots across more than 300 facilities in 2025; DHL and Locus reported 500 million picks across more than 35 DHL-managed sites in 2024. These figures establish scale, not independent ROI. [5] [6]
Measured category signals
Logistics leads professional service robotics.
102,900Transportation and logistics robots sold2024 sample · +14% · IFR [1]
+31%Professional RaaS fleet growth>24,500 units in 2024 · IFR [2]
Relevant demand signal
Transportation and logistics represented more than half of reported professional service-robot sales.
IFR recorded almost 200,000 professional service robots sold in 2024, up 9%; its logistics figure was 102,900 units. Both are survey-sample results, not whole-market estimates. [1]
Platform migration
ROS 1 support ended on 31 May 2025.
The ROS project states that Noetic—the final ROS 1 release—no longer receives features, security updates, bug fixes, or updated binaries. We infer that migration work creates an integration window for modular validation interfaces. [7]
Go-to-market hypothesis
Start where mobile autonomy is already shipping.
Initial customer hypothesis · Goodness strategy
Logistics AMR developers and integrators moving from simulation into controlled pilots—not a claimed market-size estimate.
Reachable ecosystem—not SOM
A concentrated supplier and partner landscape.
IFR tracks 333 logistics-robot suppliers and 944 service-robot producers, 80% of them SMEs. These counts describe producers—not buyers, revenue, TAM, or SOM. [2]
Current evidence wedge
Demonstrate repeatability, then audit robustness.
The Andino study separates 100/100 manually observed completions from 96/100 zero-overlap results in unobserved varied-condition simulation. It demonstrates the validation method—not physical-hardware safety. Andino evidence ↑
Lower-friction entry hypothesis
Validate the command path without replacing the robot.
A simulation-first, platform-neutral layer can enter through ROS 2 interfaces and existing test workflows before a controlled hardware pilot. Customer discovery must still validate willingness to adopt and pay.
Technology value snapshot
What is demonstrated—and what still has to be priced.
1,000 randomized benchmark trials plus 200 Andino trials across observed and varied conditions.
Fewer collision-runs than the basic controller in the best tested Mavic mode. See benchmark ↑
Every manually observed Andino run reached its target in the recorded condition.
Varied-condition Andino runs reaching target with zero solid-footprint overlap. See evidence ↑
Investor read
The evidence reduces technical uncertainty. It does not establish a company valuation.
A defensible dollar value still requires IP diligence, physical-pilot results, customer demand, pricing, revenue or contracted pilots, margins, and financing terms. The next value inflection is a controlled hardware pilot, followed by a paid design partner and repeatable deployment economics.
Plain English: more mobile robots are entering real workflows; failures become more consequential at scale; Goodness is building a check between an autonomous decision and physical motion. The current proof is measurable simulation—not yet commercial or hardware proof.
BeachheadLogistics AMR pilots→
ExpandMulti-site fleets→
BroadenAdjacent mobile robots→
PlatformAutonomous decision validation
Research audit · 23 Aug 2026: IFR service-robot figures come from a 294-supplier sample and are not projected to the whole industry; its producer counts exclude system integrators. Amazon and DHL figures are company-reported deployment milestones, not independent ROI studies. The customer profile and expansion path are Goodness hypotheses—not third-party forecasts. The EU summary is not legal advice; applicability depends on the product, role, and use case. Andino evidence is controlled Windows Webots simulation, not hardware certification.
Contact
Ready for the deeper walkthrough?
Pick a time and we will show the robot demo, benchmark comparison, current limits, validation positioning, and next pilot steps.
Book a Private Walkthrough
No form. Just choose a time.