Human-Robot Interaction: Research Landscape 2025-2026
Sources: Research synthesis, 2026-03-22; Session notes, 2026-04-08; Habibian et al. IJRR 2024; Wang et al. CHI 2026; Heron & Lau (Laurentian), arXiv:2603.00154, 2026-03; Yakolli et al., arXiv:2512.00027, 2025-11 Raw: Human-Robot Interaction: Current Research Landscape; MIT VR Haptics HCI HRI Research Groups; [Communicating Robot Learning](../../raw/clippings/Communicating Robot Learning During HRI.md); [HRI in the Age of LLMs](../../raw/clippings/HRI in the Age of LLMs Systematic Review.md); Trust in Autonomous HRC; [VLN for HRC Survey](../../raw/clippings/Vision Language Navigation Human Robot Collaboration Survey.md) Updated: 2026-08-12
Overview
Human-Robot Interaction in 2025-2026 is splitting along two axes that are developing somewhat independently: Physical HRI (cobots, surgical robots, dexterous manipulation — primarily a control and safety problem) and Social HRI (companion robots, assistive systems, conversational agents — primarily a perception and interaction design problem). LLMs are the biggest recent inflection point: before them, robots had scripted interaction patterns. After them, robots can understand context, generate situational dialogue, and reason about language-described tasks.
Physical HRI
Physical HRI is about safe co-presence and co-manipulation. The dominant standard is ISO/TS 15066, which specifies force and pressure limits for contact between humans and collaborative robots (cobots) in manufacturing.
The standard is contested: the 2026 critical review argues that energy transfer — not instantaneous force or pressure — is the correct safety metric. Current design approximations introduce unquantified performance penalties. The 2025 revision of ISO 10218-2 reflects a deeper shift: safety is no longer a property of the robot, but a property of the application. This has practical consequences for how robots are certified.
Dexterous manipulation is beyond the standard’s scope: ISO/TS 15066 was designed for cobot assembly lines. 21+ DOF dexterous hands operating in contact-rich, unpredictable scenarios exceed what the standard was designed to handle. Safety verification frameworks for these systems don’t yet exist.
Social HRI
Social HRI concerns how robots perceive human state and generate socially appropriate responses. LLMs are the dominant enabling technology for this, replacing scripted interaction patterns with contextual sensing, generative dialogue, and common-sense reasoning.
The CHI 2026 systematic review (86 papers) documents that LLM adoption in HRI is real and broad — but identifies systematic gaps:
- Research is fragmented; no standardized evaluation metrics
- Overinvestment in technical capability, underinvestment in human-centered factors (user modeling, appropriate autonomy, trust)
- Almost no longitudinal studies — human adaptation to robots over time is not studied
Trust is undertheorized: trust appears as a theme in nearly every HRI survey but is rarely operationally defined. Fear responses (especially in older adults) span 7 dimensions including the Uncanny Valley effect, privacy concerns, and technology unfamiliarity. The field lacks a unified model of how trust develops and breaks down.
Responsive interaction as a trust lever (Heron & Lau, arXiv:2603.00154, pilot study N=24): affect-responsive robots (adapting based on inferred interaction state) yielded substantially higher post-interaction trust than neutral reactive robots: ~26 points higher on the TI-HRC scale, ~15 points higher on the TPS-HRI scale — independent of overall task accuracy. Key finding: “trust emerges from interaction process dynamics rather than task performance alone.” Interaction quality (encouragement, empathy) drove the effect; communication breakdown (speech recognition failure) attenuated it. Caveat: pilot study, small N, single robot platform (Misty-II).
Multimodal Perception
The technical frontier bridging physical and social HRI. Robots that can only see or only hear are fundamentally limited. The 20-year retrospective (Frontiers 2025) shows that each modality (vision, audio, language, touch) was studied in isolation for decades; integration is the current open problem.
Generalization from lab to real deployment remains unresolved: models trained in controlled lab settings fail in real homes, hospitals, and factories. The root cause is distribution shift between training data and deployment environment — the same failure mode that plagues VLA simulation-to-real transfer.
The Bidirectionality Gap
All major HRI surveys identify the same structural gap from different angles: most current HRI systems are command-execution pipelines. Humans instruct; robots execute. True interaction requires bidirectionality:
- Robots that ask clarifying questions
- Robots that signal uncertainty
- Robots that negotiate task completion
- Robots that maintain dialogue coherence across multiple turns
The VLN survey (Yakolli et al., ~200 papers surveyed) identifies this as a critical missing capability. No deployed system currently achieves it. Building it requires: uncertainty quantification (when does the robot know it doesn’t know?), proactive communication design (how does the robot ask without being annoying?), and dialogue turn-taking across multiple exchanges.
The SIA Framework
The CHI 2026 review proposes Sense-Interaction-Alignment (SIA) to replace the classical Sense-Plan-Act model:
- Sense: contextual sensing from language — understanding situational context, not just detecting objects
- Interaction: generative socially-grounded dialogue rather than scripted responses
- Alignment: continuous adaptation to user mental model and preferences
LLMs deliver the Sense and Interaction layers. The Alignment layer — actually adapting to individual users over time — remains the research gap.
VR Haptics and HCI: Key Research Groups (MIT)
A landscape scan of MIT CSAIL and Media Lab research groups relevant to VR haptics and HRI (source: 2026-04-08):
VR Haptics — most direct:
- Hiroshi Ishii (Tangible Media Group, Media Lab): VR Haptics at Home and related work in tangible interfaces
- Stefanie Mueller (HCI Engineering, CSAIL): Haptic Links, CLAW, and fabrication-oriented haptic devices — engineering/hardware emphasis
Cross-department umbrella:
- MIT Center for Advanced Virtuality: cross-school (not CSAIL or Media Lab) VR research umbrella unit
HRI core:
- Julie Shah (Interactive Robotics Group, CSAIL/AeroAstro): task planning, human-robot teaming — physical HRI emphasis
- Cynthia Breazeal (Personal Robots, Media Lab): social robot platforms, long-term human-robot relationships — social HRI emphasis
CSAIL vs. Media Lab distinction in this space: CSAIL haptics research skews fabrication and interaction hardware (what can be engineered); Media Lab skews philosophy and experience design (what should tangibility mean). Both are relevant for LfD data glove research — CSAIL for engineering the device, Media Lab for designing the human experience around wearing it.
Adjacent (wearable sensors, not haptic output): Pattie Maes (Fluid Interfaces, Media Lab) — wearable cognitive augmentation; overlaps with glove sensor design at the wearable sensing layer.
Robot-to-Human Communication During Learning
Communicating Robot Learning (Habibian et al., IJRR 2024, arXiv:2312.00948): Most LfD research treats the human as a passive data source. This review asks: what if robots communicate what they’ve learned back to the human? The answer: it significantly changes how humans teach.
Communication interface taxonomy:
- Visual: visualizations of learned reward functions, uncertainty, intended trajectories — most common but lowest bandwidth for internal state
- Haptic: force-reflecting teleoperation — high bandwidth for physical tasks, uniquely valuable for manipulation; underexplored relative to visual
- Auditory: alerts, sonification of robot confidence — useful for attention management; rarely used alone
Key finding: when robots communicate their learning state (not just execute), humans adapt their teaching strategy. Closed-loop co-adaptation leads to better demonstrations, increased trust, and faster skill acquisition.
Haptic feedback is uniquely valuable for kinesthetic teaching because it is in-band with the demonstration interface itself — the human feels robot resistance without splitting attention to a separate display. This is a practical argument for force-reflecting teleoperation over vision-only feedback in LfD.
Connection to HITL: this work provides the theoretical grounding for why HITL (DAgger, RoboCopilot) works — it is not just about correcting errors, but about a bidirectional channel that improves human demonstration quality.
Open Questions
- How should safety standards evolve for 21+ DOF dexterous robots operating in contact-rich, unpredictable scenarios?
- What is the right autonomy level for LLM-driven robots? When should a robot defer to a human vs. act autonomously?
- How do we build evaluation benchmarks for social HRI that measure human experience, not just task completion?
- What does human-robot collaboration look like after 6 months of daily use? Trust and mental model formation over time are unexplored — no longitudinal studies exist.
- How do cultural differences in human-robot social norms affect global deployment of the same robot system?