💻 computer science

Humanoid Factors: Design Principles for AI Humanoids in Human Worlds

This paper introduces "humanoid factors" as a four-pillar framework (physical, cognitive, social, and ethical) to guide the design, evaluation, and governance of AI humanoids, ensuring they can effectively and safely coexist and collaborate with humans in shared environments.

Original authors: Xinyuan Liu, Eren Sadikoglu, Ransalu Senanayake, Lixiao Huang

Published 2026-04-17

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Xinyuan Liu, Eren Sadikoglu, Ransalu Senanayake, Lixiao Huang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are building a new kind of roommate. This roommate isn't a human, but a robot that looks like a human, walks like a human, and can talk like a human. You want to put this robot in your home, your office, or a hospital.

For decades, engineers designed tools (like hammers or cars) to fit us. They asked, "How do we make this tool safe and easy for a human to use?" This is called Human Factors.

But now, we have Humanoid Robots. They aren't just tools; they are "agents" that live in our world. The old rules don't work anymore. If you treat a robot like a hammer, you'll get hurt. If you treat it like a human, you'll be disappointed.

This paper introduces a new rulebook called Humanoid Factors (HoF). Think of it as a "Dual-Agent Design Manual." It's a guide to making sure humans and humanoids can live together without crashing, confusing, or annoying each other.

Here is the breakdown of their four main pillars, using simple analogies:

1. The Physical Pillar: "The Body Suit"

The Problem: A robot might be strong enough to lift a box, but if it moves too fast or jerks around, it looks scary. Or, it might run out of battery after 10 minutes.
The Analogy: Imagine a bodybuilder wearing a suit made of glass. They are strong, but if they move too quickly, the suit shatters.
The HoF Fix: We need to design the robot's body so it fits our world (doors, stairs, chairs) and moves in a way that feels "safe" to us. It's not just about can it do the task, but how it does it. Does it move smoothly like a dancer, or jerkily like a glitching video game character?

2. The Cognitive Pillar: "The Brain & The Brake"

The Problem: A robot might be smart enough to solve a puzzle, but if it takes 10 seconds to decide to pick up a cup, you'll think it's broken. Or, if it's unsure, it might just freeze.
The Analogy: Think of a self-driving car. It needs to know what is happening (a ball rolling into the street), what to do next (brake), and when to stop. But it also needs a "Brake" that says, "Wait, I'm not sure, let me ask a human."
The HoF Fix: The robot needs a "brain" that doesn't just calculate the fastest path, but the most predictable path. It needs to signal its thoughts (e.g., "I'm thinking about that") so you aren't surprised. It also needs "guardrails" to stop itself from doing something dangerous if it gets confused.

3. The Social Pillar: "The Dance Partner"

The Problem: Humans have unspoken rules. We don't stand too close, we wait our turn to talk, and we look at each other when we speak. If a robot ignores these rules, it feels creepy or rude.
The Analogy: Imagine dancing with a partner who never listens to the music, steps on your toes, and stares at you without blinking. You'd stop dancing.
The HoF Fix: The robot needs to learn the "dance steps" of human society. It needs to guess what you are thinking (e.g., "They look stressed, I should be quiet") and adjust its behavior. It's about calibrating trust: looking like you are smart enough to help, but not so smart that you trick us.

4. The Ethical Pillar: "The Rulebook & The Mirror"

The Problem: Who is responsible if the robot breaks your vase? Does it remember your private conversations? Does it treat everyone fairly, or does it ignore people with accents?
The Analogy: Imagine a robot butler who secretly records your family arguments to sell to advertisers, or one that refuses to help a specific person because of their skin color.
The HoF Fix: We need to build "conscience" into the code. The robot must know what data it is allowed to keep, who it can hurt (even accidentally), and how to admit when it makes a mistake. It's about making sure the robot is a good citizen, not just a good worker.

The "Fitts' Law" Experiment: Why "Success" Isn't Enough

The authors did a cool experiment to prove their point. They taught a robot to point at a target.

The Old Way: Did the robot hit the target? Yes. Great job!
The Humanoid Factors Way: How did it hit the target?

They found that the robot could hit the target, but it moved in a weird, robotic way that didn't match how humans move. Humans naturally slow down when a target is small or far away (a rule called Fitts' Law). The robot didn't do this.

The Result: Even though the robot "succeeded" at the task, a human standing next to it felt uneasy because the robot's movement was unpredictable. It was like a driver who hits the red light perfectly but drives at 100mph the whole time. It's "correct," but it's not safe or comfortable.

The Big Picture: Why This Matters

The paper argues that we are moving from an era where robots are tools (like a drill) to an era where they are teammates (like a coworker).

For Designers: You can't just build a fast robot. You have to build a robot that fits into the messy, emotional, and physical reality of human life.
For Investors: A robot that is fast but scary will fail in the market. A robot that is slightly slower but feels "natural" will win.
For Us: We need to understand that these robots aren't humans. They are machines that mimic humans. We need to teach them our rules, and we need to teach ourselves how to live with them.

In short: The paper says, "Don't just ask if the robot can do the job. Ask if the robot can do the job without making us feel like we're living in a horror movie."

1. Problem Statement

The rapid advancement of AI foundation models has enabled the development of general-purpose humanoid robots capable of operating in unstructured human environments (homes, workplaces, public spaces). However, current design and evaluation paradigms remain rooted in Human Factors (HF), which optimizes systems for human performance alone, treating robots as passive tools.

This approach is insufficient for humanoids because:

Dual-Agent Coexistence: Humanoids are no longer just tools; they are active, intelligent agents that must coexist and collaborate with humans.
Expectation Mismatch: Humanoids possess human-like morphology, leading humans to attribute human-like intent, social competence, and agency to them. When a robot's actual capabilities (driven by AI) do not match these social expectations, it leads to trust erosion, misuse, or safety risks.
Fragmented Literature: Research is currently siloed across robotics, machine learning, and ergonomics, lacking a unified framework to address the physical, cognitive, social, and ethical challenges of human-humanoid coexistence.
Evaluation Gaps: Current metrics focus heavily on task completion (e.g., "did the robot pick up the object?") while ignoring how the robot behaves (legibility, timing, social comfort), which are critical for human acceptance and safety.

2. Methodology: The Humanoid Factors (HoF) Framework

The authors propose Humanoid Factors (HoF) as a complementary framework to traditional Human Factors. HoF is designed to specify the design, evaluation, and adaptation requirements for humanoid performance within human-built spaces.

The framework is structured around four interdependent pillars, each containing specific layers:

A. Physical Pillar (P)

Focuses on the robot's embodiment and interaction with the physical world.

P1 Morphology: Defines body plans (e.g., M-Static, M-Wheel, M-Biped, M-Full) and their compatibility with human-scale affordances (doors, stairs).
P2 Sensing: Requires egocentric vision, active gaze control, and multimodal sensing (proprioception, tactile, audio) to match human perceptual capabilities.
P3 Action: Covers actuation and expressivity (how the robot communicates intent through motion, gaze, and timing).
P4 Operation: Addresses energy-compute-thermal constraints, robustness, fall resilience, and serviceability (MTBF/MTTR).

B. Cognitive Pillar (C)

Focuses on the robot's internal reasoning and decision-making processes.

C1 Situational Awareness: Perception, comprehension, and projection (predicting future states).
C2 Reasoning, Planning, & Execution: Separation of reasoning (meta-cognition), planning (goal decomposition), and execution (low-level control).
C3 Cognitive Guardrails: Self-regulation mechanisms to ensure safety, legibility, and ethical boundaries (e.g., slowing down when uncertain).
C4 Memory & Cognitive Load: Managing working, semantic, and episodic memory while balancing computational resources and minimizing human cognitive load.

C. Social Pillar (S)

Focuses on the robot's ability to interpret and influence human behavior.

S1 Inference: Estimating human internal states (intent, emotion, workload) using a "Latent Theory of Mind."
S2 Prediction: Anticipating human actions and interaction trajectories.
S3 Intervention: Determining how and when to intervene (explicitly via language or implicitly via motion/timing).
S4 Social Appearance: Aligning the robot's physical look with its actual capabilities to manage user expectations and prevent the "uncanny valley."

D. Ethical Pillar (E)

Focuses on normative constraints and accountability.

E1 Safety & Nonmaleficence: Minimizing physical and psychological harm; preserving human dignity.
E2 Privacy & Data Ethics: Data minimization, retention policies, and the "right to be forgotten."
E3 Governance: Mechanisms for accountability, incident reporting, and protection against vandalism.
E4 Fairness & Social Equity: Mitigating algorithmic bias and ensuring equitable treatment across demographics.

3. Enabling Technology: AI Foundation Models

The paper analyzes how AI Foundation Models (Generative AI, Vision-Language-Action models) enable the HoF framework.

Training Stages: The authors propose a three-stage training pipeline:
1. Pre-training: Learning general world knowledge and embodiment priors from internet-scale data.
2. Mid-training: Domain shaping using structured data (reasoning traces, physics simulations) to improve planning and robustness.
3. Post-training: Human alignment via preference learning (RLHF/DPO) to calibrate behavior, safety, and social norms.
Data Regimes: Emphasizes the need for diverse data sources, including human activity videos, synthetic data for edge cases, and human preference signals to align with HoF pillars.
Limitations: The authors warn against over-attributing agency to models. Foundation models provide capabilities but lack intrinsic goals; they require external guardrails and system-level integration.

4. Experimental Results: Illustrative Case Study

To demonstrate the utility of HoF, the authors conducted an experiment evaluating a Behavior Cloning (BC) policy on a Unitree G1 humanoid robot performing a pointing task.

Hypothesis: A robot trained solely for geometric task completion would fail to exhibit human-like temporal scaling (speed-accuracy trade-offs) predicted by Fitts' Law.
Method:
- Humans demonstrated pointing tasks at varying distances and target widths.
- A BC policy was trained to mimic these movements.
- Performance was evaluated in simulation and on the real robot.
Findings:
- Human Baseline: Strong adherence to Fitts' Law ( $R^2 = 0.74$ ), showing natural speed-accuracy trade-offs.
- Simulation: The policy showed a parallel trend ( $R^2 = 0.60$ ), suggesting partial cognitive alignment.
- Real-World Deployment: The relationship collapsed ( $R^2 \approx 0.00$ ). Physical constraints (gravity, joint compliance, friction) caused the robot to drift or oscillate, breaking the cognitive plan.
Conclusion: Standard task-completion metrics would have deemed the robot "successful" (it reached the target). However, HoF evaluation revealed a Cognitive-Physical misalignment: the robot's motion was unpredictable and uncomfortable for human partners, highlighting the failure of task-centric metrics to capture human compatibility.

5. Key Contributions

Theoretical Framework: Introduction of Humanoid Factors (HoF) as a distinct, four-pillar framework (Physical, Cognitive, Social, Ethical) specifically for designing and evaluating humanoids as co-existing agents.
Operationalization: Mapping HoF requirements to AI foundation model training pipelines (pre-training, mid-training, post-training) and data regimes.
Evaluation Shift: Proposing a shift from "task completion" to "human compatibility" metrics. The paper demonstrates using human-derived psychophysical laws (Fitts' Law) as diagnostic primitives to expose deployment failures.
Stakeholder Analysis: Providing a comprehensive framework (Table 6) for how different stakeholders (engineers, policymakers, users, insurers) interact with each HoF pillar.
Critical Limitations: Explicitly acknowledging risks such as anthropomorphic overreach, the limited transferability of human theories to machines, and the danger of premature standardization.

6. Significance and Implications

For Researchers: Moves the field beyond isolated robotics or AI benchmarks toward dual-agent co-design. It calls for shared experimental primitives that measure legibility, trust calibration, and cognitive load.
For Industry & Investors: Reframes "deployability" as a function of social and ethical alignment, not just technical dexterity. HoF metrics (e.g., intervention frequency, trust decay) are presented as leading indicators of commercial risk.
For Policy & Society: Highlights the need for new standards that address the unique challenges of adaptive, learning-based humanoids. It advocates for "humanoid literacy" and public education to manage expectations and ensure safe coexistence.
Overall Impact: The paper argues that the future success of humanoids depends not on making machines human-like, but on making them human-compatible. The HoF framework provides the necessary lens to achieve this by integrating engineering, psychology, ethics, and social science into a unified design philosophy.