Human-Robot Interaction (HRI) Deep Knowledge Pack

HRI=field studying interactions b/w humans & robots; integrates robotics, HCI, psych, soc, design. Core goal: enable safe, intuitive, effective collaboration. Key components: perception (multimodal sensors: cam, mic, LiDAR), cognition (intent recognition, ToM, context-awareness), action (motion planning, social navigation), communication (verbal, nonverbal, GUIs). Modalities: speech (ASR, TTS), gestures (pose est, sign lang rec), gaze (eye-track), haptics (force feedback), proxemics (spatial behavior). Social robotics principles: engagement, turn-taking, expressiveness, rapport. HRI frameworks: JND (Joint Negotiation & Dialogue), CWA (Cognitive Work Analysis), SLM (Social Level Model). Eval metrics: task perf, efficiency, user sat (NASA-TLX), trust (SUR, Godspeed), safety (ISO 10218-1, ISO/TS 15066). Domains: industrial (cobots, e.g., UR, Fanuc CRX), healthcare (rehab, surgery, companion bots), education (tutoring robots, e.g., NAO), service (delivery, concierge, e.g., Tally, Pepper), military (UGVs, bomb disposal). Cobot safety: speed/separation monitoring, power/force limiting (PFL), collaborative zones. ISO/TS 15066 defines PFL thresholds: 140N force, 100J energy, limb-specific limits. Trust calibration: over-trust → complacency; under-trust → disuse. Mitigation: transparency (explainable actions), feedback (auditory, visual cues), adaptive autonomy. ToM (Theory of Mind): robot models human beliefs/intents; implemented via Bayesian models, I-POMDPs. Intent prediction: LSTM, Transformer-based seq models on traj data. Shared control: blending human input (joystick, EMG) w/ autonomy; impedance ctrl for smooth transitions. Communication latency: >200ms disrupts turn-taking; real-time systems target <100ms. Multimodal fusion: late fusion (score-level), early fusion (raw data), model-based (graph nets). Failures: miscommunication, unexpected motion, social norm violation. Recovery: graceful degradation, clarification queries, fallback behaviors. Ethical concerns: privacy (sensor data), consent (autonomy), bias (training data), deception (anthropomorphism). Bias mitigation: diverse datasets, fairness-aware ML. Long-term interaction: user adaptation, personalization (user models, preferences), memory retention. Learning from Demonstration (LfD): kinesthetic teaching, teleop → policy extraction via IRL, DMPs. Explainable AI (XAI): saliency maps, natural lang explanations, counterfactuals. Current SOTA: models like PaLM-E, RT-2 integrate vision-lang-action; end-to-end VLA (Vision-Language-Action) policies. Social nav: STORM, TFPP models predict human traj; use social forces, GANs. Emotion rec: multimodal (facial, speech, physiological); FER (Facial Emotion Recognition) via CNNs; valence/arousal est. Affective computing: emotion synthesis via facial expr (FACS), voice modulation. ROS-based HRI: ROS2 (DDS, real-time), SMACH (state machines), MoveIt (motion), Gazebo (sim). Tools: PyRobot, HRI.py, OpenFace, MediaPipe. Common pitfalls: over-anthropomorphizing robot capab, poor user mental model, ignoring cultural norms, insufficient testing w/ diverse users. Design guidelines: user-centered design, iterative prototyping, inclusive testing. Future directions: neuroadaptive interfaces (EEG-fMRI), large behavior models (LBMs), ethical-by-design frameworks, long-horizon autonomy. Standards: IEEE P7008 (ethical autonomous systems), ISO 13482 (personal care robots). Evaluation: within-subjects vs between-subjects designs, longitudinal studies, Wizard-of-Oz (WoZ) for prototyping. WoZ limitations: operator bias, inconsistent behavior. Metrics: SUS (System Usability Scale), UMIS (User Experience Questionnaire), trust surveys. Case study: Da Vinci surgical system — teleop, haptic feedback (limited), 3D vision; challenges: latency, cost. Case study: Boston Dynamics Atlas — dynamic motion, but limited HRI; focus on autonomy over interaction. Emerging: swarm HRI (multi-robot, single human), meta-learning for rapid personalization, embodied LLMs (e.g., Google's RT-2, NVIDIA's VIMA). Critical challenge: alignment — ensuring robot goals/values match human expectations. Research gaps: cross-cultural HRI, aging populations, scalable personalization, real-world robustness.