CRISP: How Humanoid Robots Are Learning to Criticize and Refine Their Own Social Behavior

What if a robot could watch itself interact with a human, decide its own behavior was awkward, and silently fix it — without anyone telling it to? That's the core idea behind CRISP, a new framework accepted to ICRA 2026 that turns a Vision-Language Model into an autonomous inner critic for humanoid social behavior.

The Problem With Scripted Sociability

Humanoid robots entering human environments face a challenge that goes beyond locomotion and manipulation: they need to behave socially. A wave, a nod, an appropriate gesture during a handshake — these small signals matter enormously to the humans sharing space with them. Yet most robots today handle these situations through predefined motion libraries or by relying on human operators to supply and tune behaviors. That approach is brittle. It doesn't scale. And it puts a permanent human in the loop for every new scenario the robot encounters.

Researchers Jiyu Lim, Youngwoo Yoon, and Kwanghyun Park from KwangWoon University and ETRI (Electronics and Telecommunications Research Institute) in South Korea set out to break this dependency. Their answer is CRISP — Critique-and-Replan for Interactive Social Presence — a framework accepted to the 2026 IEEE International Conference on Robotics and Automation (ICRA) in Vienna, Austria.

An Inner Critic Built From a VLM

The central insight of CRISP is deceptively simple: if a Vision-Language Model can understand images and language well enough to describe human social scenes, it can also evaluate whether a robot's behavior fits into one. CRISP deploys a VLM not just as a planner, but as a social critic — a module that watches rendered visualizations of the robot's proposed motions and scores how socially appropriate they look.

This makes the system self-contained. The robot doesn't need a human observer or a specialized reward model trained on robotics data. It just needs the same general-purpose VLM that can already reason about social context from visual input. The robot, in effect, learns to see itself through human eyes.

The Five-Step CRISP Pipeline

CRISP operates as a structured five-stage loop that takes a situation description all the way to polished, socially appropriate joint-level motion:

1. Structural extraction: The system reads the robot's description file (e.g., an MJCF file) to identify all movable joints and their physical constraints — no robot-specific API required.
2. Behavior planning: Given a situational context (e.g., "greet a visitor at the door"), the VLM generates a step-by-step natural language plan for what the robot should do.
3. Low-level code generation: The VLM translates that plan into joint control code, guided by visual renderings of each joint's range of motion to keep movements physically plausible.
4. Social critique: The VLM evaluates the resulting motion for social appropriateness, assigning a score out of 10 and identifying exactly which steps fall short.
5. Iterative refinement: Steps 3 and 4 repeat in a reward-based search loop, refining the motion until it scores 8 or higher. Only then is the behavior accepted.

The threshold of 8/10 is a deliberate design choice — demanding enough to filter out awkward behaviors, lenient enough to avoid infinite loops on genuinely ambiguous social scenarios.

Validated Across Robots, Proven With People

The team tested CRISP across 20 diverse social scenarios on five different robot platforms, spanning humanoids and mobile manipulators. Because the framework only requires the robot's structural file, it transferred across platforms without retraining or platform-specific engineering — a meaningful result for a field where generalization is hard-won.

More compelling is the human subject study. Fifty participants evaluated robot behaviors generated by CRISP versus GenEM, a leading baseline for generative embodied motion. CRISP scored 4.5 ± 1.11 on a 5-point social appropriateness scale against GenEM's 3.4 ± 1.13 — a statistically significant improvement with p < 0.001. Human observers consistently judged CRISP-generated behaviors as more natural and contextually fitting.

"The key insight is that robots can self-assess social appropriateness without explicit human feedback — just by 'watching' themselves through the VLM's visual understanding."

— Lim et al., CRISP: Critique-and-Replan for Interactive Social Presence, ICRA 2026

Why This Matters for Humanoid Deployment

As humanoid robots move from controlled labs into homes, offices, and public spaces, social competence is becoming a real deployment bottleneck. A robot that can navigate a warehouse but unnerves every human it passes isn't production-ready. CRISP addresses this not by hardcoding more behaviors, but by giving robots the machinery to develop and refine appropriate behaviors autonomously — from a one-time structural description of who the robot is, and ongoing VLM-guided introspection about how it's behaving.

The broader implication is significant: social intelligence may increasingly come from the same foundation models already driving language and vision tasks. CRISP is an early, concrete demonstration that this transfer is not just theoretically attractive — it's experimentally validated and practically deployable across different robot form factors today.