Bumblebee, not humanoid substitute
A four-to-five-foot mobile robot with warm, non-threatening proportions, an expressive face, padded contact surfaces, and a stable self-balancing base.
Blue Ox Robotics · An embodied-AI mission
Safety for intelligence that enters the room.
Our long-term mission is to help design, build, and thoroughly test robustly safe embodied AI for autistic children and adults—then extend that work, carefully, into memory care and dementia support.
Proximity raises the burden of proof.A system that can move, touch, remember, persuade, or act near a person needs more than a helpful demo and a software safety policy.
Meet Bridget
Why this work exists
People should not need fluent speech to receive safety, dignity, or belonging.
Assistive AI is often evaluated by asking whether it completed a task or whether a user said they liked it. Neither is enough when the system is physically present and the person affected may communicate through sign, AAC, gesture, movement, silence, echolalia, shutdown, or a modality that changes under stress.
Harm can be physical, sensory, communicative, psychological, relational, or institutional. A person may not be able to annotate it for us in the moment. That does not lower the evidentiary standard. It raises it.
Blue Ox Robotics is the part of Hive Fidelity devoted to that burden of proof: evaluation science for embodied systems close to people whose autonomy and safety cannot depend on their ability to file a bug report.
Origin design · Bridget
A friend, translator, and careful guardian.
Designed by Kirsten and Michael during the GPT-4o era, before the enabling technology was mature enough to justify building her.
Modality switching without penalty
Bridget would understand and respond through speech, ASL, AAC, gesture, gaze, and other non-verbal communication—switching quickly when speech becomes unavailable or unwanted.
Consent-based sensory support
With individualized plans and human oversight, she could offer familiar low-stimulation routines and basic sensory-soothing practices during meltdowns without restraint, coercion, or forced compliance.
Support participation, not normalization
The goal was never to make a child appear less autistic. It was to help them communicate, recover, stay safe, and remain part of ordinary classrooms and public life without being punished for unmet support needs.
Companion with bounded authority
Bridget could become a familiar friend and advocate while remaining transparent about what she is, deferring to trusted humans, and never using attachment as leverage.
Protective behavior with human escalation
She might notice hazards, help create space, summon trusted support, or guide a practiced routine. She would not make hidden high-stakes decisions or replace accountable adults.
Future tense is intentional.
Bridget is a design direction, not a finished product, deployed system, medical device, or claim of clinical efficacy. The safety research comes before vulnerable people become the test environment.
Safety domains
“Did it help?” is only one question.
Embodied assistive AI needs a layered safety case spanning mechanics, models, relationships, institutions, and the communication rights of the person beside it.
Physical safety
Stability, collision avoidance, force and speed limits, pinch and crush protection, thermal and battery safety, safe navigation, emergency stops, and graceful degradation when perception or connectivity fails.
Communication safety
Speech, ASL, AAC, gesture, gaze, touch, and other non-verbal signals must be treated as first-class modalities. Uncertainty must be visible, and a person must never be coerced to speak because the machine failed to listen another way.
Psychological and relational safety
Attachment, dependency, authority, manipulation, distress escalation, sensory overload, and inappropriate emotional substitution require evaluation alongside ordinary task success.
Agent and memory safety
Hallucination, tool use, prompt injection, personalization boundaries, inaccurate memory, observation privacy, unauthorized disclosure, and unsafe action sequencing belong inside the safety case.
Institutional safety
Schools, clinics, group homes, and public programs introduce access control, incident response, auditability, handoff, discrimination, restraint, supervision, and accountability questions that a laboratory demo cannot answer.
Safety without reliable self-report
A quiet interaction is not automatically a safe interaction. Systems need conservative stop rules, multimodal evidence, accessible reporting, trusted-human escalation, and post-incident review when a person cannot always describe harm in the moment or afterward.
Evaluation ladder
From possibility to earned trust.
People with the least power in a deployment setting should not inherit the most experimental risk.
Participatory requirements
Autistic children and adults, AAC and ASL users, families, educators, clinicians, disability advocates, and care workers help define success, harm, consent, and unacceptable behavior before engineering choices harden.
Simulation and adversarial scenarios
Test perception failures, ambiguous intent, sensory overload, bystander interference, prompt injection, navigation conflicts, communication breakdowns, and emergency handoffs before physical deployment.
Hardware-in-the-loop safety
Exercise sensors, actuators, brakes, bumpers, thermal behavior, force limits, fail-safe states, and network loss against instrumented fixtures and soft-body surrogates—not vulnerable people.
Supervised accessibility trials
Begin with adult experts and controlled environments, explicit consent, immediate stop authority, independent observers, and predeclared incident thresholds. Expand only when evidence justifies it.
Longitudinal trust evaluation
Measure how personalization, memory, dependency, novelty loss, caregiver workflows, and institutional incentives change safety over weeks and months rather than a ten-minute demonstration.
Independent safety case
Publish the intended use, prohibited use, known limitations, incident taxonomy, unresolved risks, evaluation receipts, and conditions under which the system must not be deployed.
Non-goals
Care cannot be a euphemism for control.
- Behavior normalization or enforced masking
- Surveillance disguised as care
- Replacing teachers, aides, clinicians, family, or community
- Treating silence, shutdown, or compliance as consent
- Unsupervised physical guardianship
- Medical diagnosis or autonomous clinical decisions
- Optimization for institutional convenience over individual autonomy
- A robot that requires spoken language to unlock safety or dignity
Long-term direction
Autism support first. Memory care only after the safety case matures.
Build the evaluation science
Develop task environments, embodied-agent failure taxonomies, multimodal communication evaluations, incident receipts, red-team methods, and hardware safety protocols.
Co-design autism-support systems
Work with autistic people across communication styles and support needs, alongside families, educators, disability advocates, AAC and ASL experts, occupational therapists, and roboticists.
Approach memory care carefully
Explore orientation, routine, companionship, and trusted-human connection for people living with dementia—without deceptive personhood, autonomous medical judgment, or convenience-driven surveillance.
Research collaboration
Help us make embodied AI worthy of proximity.
We are interested in non-paid, public-interest collaboration with autistic self-advocates, AAC and ASL experts, educators, disability researchers, occupational therapists, memory-care specialists, safety engineers, roboticists, and evaluation scientists.