Introducing M: A Modular, Modifiable Social Robot
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.
Rosa: I'm Rosa.
Dev: And I'm Dev, here with Taro.
Rosa: We're looking at the paper: "Introducing M: A Modular, Modifiable Social Robot".
Dev: Here's the gist of the paper. We present M, an open-source, low-cost social robot platform designed to reduce platform friction that slows social robotics research by making robots easier to reproduce, modify, and deploy in real-world settings.
Rosa: This platform combines a modular mechanical design, multimodal sensing, and expressive yet mechanically simple actuation architecture with a ROS2-native software package that cleanly separates perception, expression control, and data management.
Dev: Platform Design Philosophy M is designed explicitly as a modular, multimodal, and reproducible infrastructure for social robotics research.
Rosa: The design rationale emphasizes several key features to address platform friction in the field.
Paper summary: Rosa: So to recap this paper, "Introducing M: A Modular Modifiable Social Robot", they’re presenting a platform called M that aims to reduce platform friction in social robotics research by making robots easier to reproduce, modify, and deploy in the real world.
Dev: The core thesis is combining a modular mechanical design with multimodal sensing and simple but expressive actuation architecture using a ROS2 native software package that keeps perception, expression control, and data management separate.
Taro: It’s about making the whole process more collaborative across different research sites by providing shared community resources to help people replicate and adapt the robot easily.
Rosa: They focus on several key design philosophies: modularity means you can swap out elements like sensors without re-engineering everything, multimodal interaction lets you study how different sensing channels combine to shape engagement strategies, and expressive embodiment uses a limited degree of freedom focused on key gestures like gaze or body orientation.
Dev: And reproducibility is a big focus because they designed it for straightforward assembly and replication through documented designs and clear software interfaces, which supports more cumulative research.
Taro: I’m interested in the deployability part too. They built it for long-term field deployment with tool-free mechanical access for rapid maintenance, plus privacy-aware operation with options like camera covers or radar sensing.
Rosa: They also integrated containerized software for consistent runtime environments and included integrated data logging that automatically adds timestamps and session metadata, plus remote monitoring for non-intrusive health checks during autonomous operation.
Dev: The software stack is structured around three core components: multimodal perception to get data from all sensors, expression control to translate that into behaviors like motion or facial expressions, and data management for tracking sessions.
Taro: What I find interesting is how the simulation environment is designed to be interface equivalent with the physical system, which allows researchers to develop perception pipelines and behavior coordination logic in simulation without needing to change anything before deployment.
Conclusion: Rosa: Thinking about the title, "Introducing M: A Modular, Modifiable Social Robot", it really speaks to the goal of making social robotics research more accessible by focusing on building tools that don't slow things down.
Dev: I think what matters is how they achieved that by separating the system into distinct assemblies like head and body so you can modify specific parts easily without touching the whole thing.
Taro: It means researchers aren't stuck on one fixed robot design; they can explore different embodiments or sensing configurations more quickly, which supports a wider range of questions about embodiment and interaction.
Rosa: And the implication for deployment is that because it’s designed for longitudinal field use with those built-in maintenance features and privacy options, it moves social robotics research out of just the lab and into actual in-home or community settings more effectively.
Dev: Ultimately, M seems positioned as a practical foundation—a platform—that allows people to focus more on the social science aspect of robot interaction rather than getting bogged down in complex engineering headaches for every single trial.
Victor Nikhil Antony, Zhili Gong, Yoonjae Kim, Chien-Ming Huang
Department of Computer Science, Johns Hopkins University · Department of Mechanical Engineering, Rice University
cs.RO, cs.HC
Submitted: 2026-03-19
Updated: 2026-10-28
Project page: https://m-website-taupe.vercel.app
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 90/100
The gist: We present M, an open-source, low-cost social robot platform designed to reduce platform friction that slows social robotics research by making robots easier to reproduce, modify…
Key concepts
- Modularity
- M uses a modular architecture where different parts like sensors or the shell can be swapped out without needing to rebuild the entire system. This significantly reduces engineering effort when research questions change, allowing researchers to easily explore new configurations for embodiment and sensing.
- Multimodal Interaction
- The platform integrates various sensing and expressive channels into one unified structure. This enables researchers to study how different combinations of expressive behaviors influence how people engage with the robot and what assistance strategies work best.
- Expressive Embodiment
- M uses a simple mechanical design focused on key gestures like gaze and body orientation, rather than complex movements. This prioritizes meaningful social signaling while keeping the hardware mechanically straightforward to build and maintain.
- Simulation Interface Equivalence
- The simulation environment precisely mirrors the physical robot's software interfaces, including perception topics and control mechanisms. This consistency ensures that behaviors tested in simulation can be reliably transferred to the physical robot without needing major code changes.
Terminology
Summary
We present M, an open-source, low-cost social robot platform designed to reduce platform friction that slows social robotics research by making robots easier to reproduce, modify, and deploy in real-world settings. p. 0 MANUSCRIPT UNDER REVIEW 1
Introducing M: A Modular, Modifiable Social Robot
Victor Nikhil Antony1
, Zhili Gong2
, Yoonjae Kim1
and Chien-Ming Huang1
Abstract—We present M, an open-source, low-cost social robot
platform designed to reduce platform friction that slows social
robotics research by making robots easier to reproduce, modify,
and deploy in real-world settings. M combines a modular mechanical design, multimodal sensing, and expressive yet mechanically simple actuation architecture with a ROS2-native software
package that cleanly separates perception, expression control, and
data management. MANUSCRIPT UNDER REVIEW 1
Introducing M: A Modular, Modifiable Social Robot
Victor Nikhil Antony1
, Zhili Gong2
, Yoonjae Kim1
and Chien-Ming Huang1
Abstract—We present M, an open-source, low-cost social robot
platform designed to reduce platform friction that slows social
robotics research by making robots easier to reproduce, modify,
and deploy in real-world settings. M combines a modular mechanical design, multimodal sensing, and expressive yet mechanically simple actuation architecture with a ROS2-native software
package that cleanly separates perception, expression control, and
data management.Where the paper says it
Where the paper says it
Platform Design Philosophy
M is designed explicitly as a modular, multimodal, and reproducible infrastructure for social robotics research. The design rationale emphasizes several key features to address platform friction in the field. p. 0 MANUSCRIPT UNDER REVIEW 1
Introducing M: A Modular, Modifiable Social Robot
Victor Nikhil Antony1
, Zhili Gong2
, Yoonjae Kim1
and Chien-Ming Huang1
Abstract—We present M, an open-source, low-cost social robot
platform designed to reduce platform friction that slows social
robotics research by making robots easier to reproduce, modify,
and deploy in real-world settings. M combines a modular mechanical design, multimodal sensing, and expressive yet mechanically simple actuation architecture with a ROS2-native software
package that cleanly separates perception, expression control, and
data management.Where the paper says it
: Modularity, M adopts a modular architecture that different elements (e.g., sensors, shell) to be adapted without reengineering the full system. p. 1 M adopts a modular architecture that different elements (e. , sensors, shell) to be adapted without reengineering the full system. , sensors, shell) to be adapted without reengineering the full system. This reduces engineering overhead
for iteration as research questions evolve, supporting exploration of embodiment, sensing and interaction configurations.Where the paper says it
Where the paper says it
: Multimodal Interaction, M integrates various sensing and expressive channels within a unified architecture, enabling investigation of how combinations of expressive behaviors shape engagement and assistance strategies for social robots. ◦
: Expressive Embodiment, M adopts a deliberately limited-degree-of-freedom embodiment that emphasizes key gesture types (e.g., gaze, body orientation, beat gestures) while maintaining mechanical simplicity. p. 1 M
adopts a deliberately limited-degree-of-freedom embodiment
that emphasizes key gesture types (e. , gaze, body orientation,
beat gestures) while maintaining mechanical simplicity. , gaze, body orientation,
beat gestures) while maintaining mechanical simplicity. By
prioritizing critical expressive cues, M enables meaningful
social signaling without high-dimensional articulation.Where the paper says it
Where the paper says it
: Reproducibility, M is designed for straightforward assembly and replication through documented designs, clearly defined software interfaces. By making reproduction practical and supporting this process through shared community resources, M aim to enable more collaborative, multi-site, and cumulative social robotics research. p. 2 : M - SOCIAL ROBOT 3
defined software interfaces, and the use of cost-accessible
components and fabrication methods. By making reproduction
practical and supporting this process through shared community resources, M aim to enable more collaborative, multi-site,
and cumulative social robotics research.Where the paper says it
: Deployability, M is designed for longitudinal field deployment with tool-free mechanical access for rapid maintenance. This includes privacy-aware operation through optional camera covers and radar-based sensing, containerized software for consistent runtime environments, integrated data logging with automatic timestamps and session metadata, and remote monitoring for non-intrusive health checks during autonomous operation. p. 2 Deployability. M is designed for longitudinal field deployment with tool-free mechanical access for rapid maintenance,
privacy-aware operation through optional camera covers and
radar-based sensing, containerized software for consistent runtime environments, integrated data logging with automatic
timestamps and session metadata, and remote monitoring for
non-intrusive health checks during autonomous operation.Where the paper says it
Mechanical Architecture Details
The mechanical design focuses on separating functions into distinct assemblies to enhance modularity and ease of modification. The system is composed of two primary assemblies, the head and the body, each designed to house specific actuation, sensing, and computation components while enabling straightforward access. p. 2 Mechanical Design
The mechanical design of M emphasizes modularity, ease
of modification, and support for expressive motion while
maintaining mechanical simplicity and robustness. The robot
is composed of two primary assemblies, the head and the
body, each designed to house specific actuation, sensing,
and computation components while enabling straightforward
access, clean wire management, and physical customization.Where the paper says it
: Head Assembly, the head assembly houses the primary computational, audio, and thermal management components. p. 2 Head Assembly. The head assembly houses the primary
computational, audio, and thermal management components,
as well as the yaw actuator used for expressive head motion
Fig. Key components include the head module (containing compute,
audio, and display), head rotation and nodding modules, customizable face
masks, articulated arms with magnetic attachments, camera, and body rotation
module with weighted base for stability. Specifically, the head contains the head yaw
motor, an onboard single-board computer (Raspberry Pi), a ReSpeaker audio module providing both microphone and speaker
functionality, an internal cooling fan for heat dissipation,
and an external I/O power switch. These components are
mounted within an internal holder that secures both the motor
and compute hardware, providing structural stability while
maintaining accessibility for maintenance and replacement. The external head shell is divided into four primary components to support modularity and customization. The left
and right base shells enclose the internal cavity and provide
structural support. A removable face plate encloses the display
and connects the base shells, allowing rapid replacement for
aesthetic or functional customization without disassembling
the full head assembly.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
: Body Assembly, the body assembly consists of a central structural frame that supports the primary actuators and connects the head and arm subsystems. The frame houses the head pitch motor, which is mechanically coupled to the head module via a pulley-based transmission. p. 2 The body assembly consists of a central
structural frame that supports the primary actuators and connects the head and arm subsystems. The frame houses the
head pitch motor, which is mechanically coupled to the head
module via a pulley-based transmission. This configuration
enables expressive pitch motion while isolating the motor from
the head enclosure, reducing inertial load on the head and
simplifying both mechanical routing and cable management. Two articulated arms are mounted on either side of the central
frame and actuated using micro-servo motors. Each arm incorporates embedded magnets to support rapid attachment of
accessories or end-effectors, enabling physical customization
without mechanical rework. A PCA9685 motor control board
is mounted on the the central frame and interfaces with all
servo motors in the system.Where the paper says it
Where the paper says it
Where the paper says it
Software and Simulation Stack
M’s software architecture is organized around a set of core functional components with clear interfaces formalized through ROS2 message, service, and action definitions. The core architecture comprises three interacting components: (1) multimodal perception, (2) expression control, (3) data management. p. 3 These interfaces are formalized through ROS2 message,
service, and action definitions that specify data formats and
control semantics, separating what each capability provides
from how it is implemented and enabling components to be
modified or replaced without restructuring the overall system. At a high level, the core architecture comprises three
interacting components: (1) multimodal perception, (2) expression control, (3) data management.Where the paper says it
: Multimodal Perception, M’s perception packages provide access to audio, video, and other sensors for the rest of the program. p. 3 M’s perception packages provide
access to audio, video, and other sensors (e. , mmwave radar,
touch) for the rest of the program. , mmwave radar,
touch) for the rest of the program. Sensor streams are exposed
as continuous topics and processed by independent perception modules responsible for tasks such as speech activity
detection, speech recognition, visual sensing, or environmental
monitoring.Where the paper says it
Where the paper says it
: Expression Control, Expression control abstracts expressive outputs (e.g., motion, facial expressions) into timeline-based behaviors. p. 3 Expression control abstracts expressive outputs (e. , motion, facial expressions) into timelinebased behaviors. Low-level control commands (e. , motor positions) are encapsulated behind controllers that expose higherlevel animation and actuation interfaces. , motor positions) are encapsulated behind controllers that expose higherlevel animation and actuation interfaces. Expressive behaviors
are represented as parameterized sequences specifying timing,
target states, and interpolation, enabling smooth transitions,
blending, and interruption.Where the paper says it
Where the paper says it
Where the paper says it
: Data Management, M provides infrastructure for session-level tracking and multimodal data logging essential for longitudinal social robot deployment. The m-logging package supports synchronized recording of sensor streams, system events, and interaction state. p. 3 M provides infrastructure for sessionlevel tracking and multimodal data logging essential for longitudinal social robot deployment. The m-logging package supports synchronized recording of sensor streams, system events,
and interaction state, enabling researchers to monitor deployment health, validate interaction execution, and reconstruct
user experiences post-deployment.Where the paper says it
Simulation and Sim-to-Real Transfer
M includes a simulation environment designed to support rapid iteration, safe development, and reproducible transfer of interaction behaviors from simulation to physical deployment. A key design principle of the simulation environment is interface equivalence with the physical system. p. 3 Simulation
M includes a simulation environment designed to support
rapid iteration, safe development, and reproducible transfer of
interaction behaviors from simulation to physical deployment. A key design principle of the simulation environment is
interface equivalence with the physical system. A key design principle of the simulation environment is
interface equivalence with the physical system. The simulated
robot exposes the same multimodal perception topics, expression control interfaces, and data management mechanisms
as the physical robot. Hardware-specific components—such
as motors, displays, and sensors—are replaced by simulated
counterparts that implement identical ROS 2 message, service, and action interfaces. This consistency allows perception pipelines, behavior coordination logic, and expressive
behaviors to be developed and tested in simulation without
modification prior to deployment on the physical robot. Hardware-specific components—such
as motors, displays, and sensors—are replaced by simulated
counterparts that implement identical ROS 2 message, service, and action interfaces. This consistency allows perception pipelines, behavior coordination logic, and expressive
behaviors to be developed and tested in simulation without
modification prior to deployment on the physical robot.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
: Simulation Interface Equivalence, The simulated robot exposes the same multimodal perception topics, expression control interfaces, and data management mechanisms as the physical robot. p. 3 A key design principle of the simulation environment is
interface equivalence with the physical system. The simulated
robot exposes the same multimodal perception topics, expression control interfaces, and data management mechanisms
as the physical robot.Where the paper says it
Real-World Application Examples
M’s capabilities are demonstrated through various extensions and interaction templates that showcase its practical utility.
: Extensions, To demonstrate M’s the modularity and extensibility, we integrated additional sensing and actuation modalities. p. 4 DEMONSTRATING USAGE
A. Extensions
To demonstrate M’s the modularity and extensibility, we
integrated additional sensing and actuation modalities. M supports a
Frequency Modulated Continuous Wave (FMCW) radar sensor
mounted on its front plate, enabling detection of human
presence and coarse motion without capturing visual or audio
data. This privacy-preserving modality allows M to remain
aware of its social context without relying on cameras or
microphones. Capacitive touch sensors embedded beneath M’s exterior shell enable the robot to sense direct
physical contact through its surface. This capability supports
interactions in which touch functions as an intentional yet
lightweight signal such as acknowledging attention, requesting continuation, or expressing reassurance (See Fig. Vibro-Tactile Actuation. M’s shell can have linear resonant
actuator (LRA) vibration motors that provide localized vibrotactile feedback through its shell.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
: Example Interaction: One-Way Storytelling, To illustrate how rich, expressive multi-modal behaviors can be deliver with M, we include an example storytelling interaction. This interaction is implemented as an example ROS 2 package that demonstrates how expressive behaviors can be authored, scheduled, and executed on M. p. 5 For instance, when narrating a winter tale,
M displays wide-eyed wonder during magical moments, tilts
its head thoughtfully during reflective passages, and using its
arms to emphasize key story beats2
(see Fig. This storytelling interaction is implemented as an example
ROS 2 package that demonstrates how expressive behaviors
can be authored, scheduled, and executed on M. This storytelling interaction is implemented as an example
ROS 2 package that demonstrates how expressive behaviors
can be authored, scheduled, and executed on M. The story and
its behaviors are generated with an LLM-based pipeline that
decomposes the narrative into sequential chunks, each paired
with corresponding expressive cues [14].Where the paper says it
Where the paper says it
: Example Interaction: Two-Way Conversation, To illustrate two-way, socially grounded interaction, M includes an example conversational positive psychology coaching interaction. In each daily session, M engages in a targeted practice session where it asks open-ended questions about positive experiences and listens to and acknowledges their responses with empathetic body language and facial expressions. p. 5 In each daily session, M engages
in a targeted practice session (see Fig. 7): the robot asks
open-ended questions about positive experiences, listens to and
acknowledges their responses with empathetic body language
and facial expressions, and asks follow-up questions that build
on what the user has shared.Where the paper says it
Future Research Directions
The discussion points toward several critical avenues for future research in social robotics.
: Foundational Models for Social Intelligence, A key direction forward is the development of foundational models for social robot behaviors that directly integrate vision, speech, and action modalities. p. 6 Moreover, extending these systems
typically requires reengineering multi-staged agentic pipelines. An important direction forward is the development of
foundational models for social robot behaviors that directly
integrate vision, speech, and action modalities. An important direction forward is the development of
foundational models for social robot behaviors that directly
integrate vision, speech, and action modalities. Unlike visionlanguage-action (VLA) models trained on physical manipulation tasks [17], these models would focus specifically on
social interaction dynamics: interpreting affect, managing turntaking, coordinating multimodal expression, and adapting to
conversational repair.Where the paper says it
Where the paper says it
: Human-AI Co-Creation of Social Robot Interactions, M’s modular software architecture and example programs are intended to lower the barrier to engineering social robotic interactions. Human-AI Co-creation as an engineering paradigm could allow researchers, domain experts, and even end users to collaboratively specify goals, constraints, and social norms. p. 7 8 MANUSCRIPT UNDER REVIEW
specifying interaction logic, debugging failure modes, and
iterating on social behaviors. Human-AI Co-creation as an
engineering paradigm could allow researchers, domain experts,
and even end users to collaboratively specify goals, constraints,
and social norms, with AI systems assisting in designing,
generating, refining, and validating interaction policies.Where the paper says it
: Designing for No-Contact, Long-Horizon Deployment, Scaling these deployments to multi-year horizons introduces new challenges. This necessitates features beyond containerized software architectures, raising questions about remote update mechanisms, fault recovery and version control across evolving interaction designs. p. 7 Long-term studies demand systems that
can be updated, maintained, and reconfigured with minimal
physical intervention, particularly when deployed with nontechnical users or special populations [18]. This requirement
necessitates features beyond containerized software architectures, raising questions about remote update mechanisms, fault
recovery, version control and backward compatibility across
evolving interaction designs.Where the paper says it
: Supporting Meaningful Community Integration, Long-term adoption will depend on broader considerations, including alignment with existing research workflows and quality of documentation. Beyond tooling, questions of community governance and standardized evaluation practices will be critical to ensuring that M functions as a shared research infrastructure. p. 7 Long-term adoption will depend
on broader considerations, including alignment with existing
research workflows, clarity on adoption tradeoffs, quality of
documentation, and incentives for sharing interaction code,
data, and experimental artifacts. Beyond tooling, questions of
community governance, standardized evaluation practices, and
mechanisms for sharing longitudinal datasets will be critical to
ensuring that M functions as a shared research infrastructure
rather than a standalone system.Where the paper says it
Limitations and Outlook
M has limitations that represent opportunities for community development. Specifically, the paper notes that long-term durability beyond week-long deployments remains uncharacterized. p. 7 LIMITATIONS
M has limitations that represent opportunities for community development. Long-term durability beyond week-long
deployments remains uncharacterized. M’s base capabilities do not include advanced
social perception or intelligence modules (e. , emotion recognition, gaze tracking, engagement estimation); We view M as
a collaborative research infrastructure that can evolve through
shared extensions and improvements, collectively advancing
the field’s capacity for longitudinal, reproducible research. M’s base capabilities do not include advanced
social perception or intelligence modules (e. , emotion recognition, gaze tracking, engagement estimation); We view M as
a collaborative research infrastructure that can evolve through
shared extensions and improvements, collectively advancing
the field’s capacity for longitudinal, reproducible research.Where the paper says it
Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Modular hardware support allows researchers to
explore embodiment, sensing and interaction configurations
withoutreengineering the full system.
This enables rapid iteration of robot form factors and sensor suites for testing social behaviors. -
ROS 2 native software package cleanly separates
perception, expression control, and data management,
allowing components to be modified or replaced without restructuring the overall system. This facilitates experimentation with alternative algorithms or sensing modalities in perception modules independently of expression control logic. -
Simulation environment with
interface equivalence to hardware
supports a workflow where interaction behaviors are developed and tested in simulationprior to deployment on the physical robot.
This reduces engineering overhead and improves safety during development by minimizing divergence between simulation and deployment. -
Integration of non-intrusive sensing, such as FMCW radar for human presence detection, allows the robot to
remain aware of its social context without relying on cameras or microphones.
The improved system can proactively engage users withsubtle, playful behaviors
upon detecting presence in privacy-sensitive contexts. -
Implementation of capacitive touch sensing enables interactions where
touch functions as an intentional yet lightweight signal such as acknowledging attention, requesting continuation, or expressing reassurance.
This allows for embodied exchanges without introducing additional visible instrumentation or altering the robot's external form. -
LLM-based pipelines generate interaction cues that specify
gesture animations and timing information relative to spoken audio,
allowing expressions to be synchronized with narrationwithout embedding low-level control logic into the interaction code.
This enables authors to rapidly prototype complex, synchronized multi-modal storytelling interactions. -
The two-way conversation example uses a state tracker to maintain context across dialogue history and session state, which then produces responses by selecting
facial expressions and body gestures using a large language model.
This enables the AI system to providecontextappropriate utterances
paired with embodied expression that adapts based on user input. -
Foundational models for social robot behaviors could be developed that
directly integrate vision, speech, and action modalities
to reason over raw signals. This would produce interactions with greater coherence between perception and expression, resulting in interactions that aremore fluid, adaptive, and context-aware.
Abstract
We present M, an open-source, low-cost tabletop social robot platform aimed at lowering barriers to social robotics research by making robots easier to reproduce, modify, and deploy in real-world settings. M combines a modular mechanical design, multimodal sensing, and expressive yet mechanically simple actuation architecture with a ROS2-native software package that cleanly separates perception, expression control, and data management. The platform includes a simulation environment with interface equivalence to hardware to support easy sim-to-real transfer of interaction behaviors. We demonstrate extensibility through additional sensing/actuation modules and provide example interaction templates for storytelling and two-way conversational coaching. Lastly, we report real-world use of M in interaction design and week-long in-home deployments, showing how M can serve as a practical platform for longitudinal, reproducible social robotics research.
Sources
- Designing Social Robots that Engage Older Adults in Exercise: A Case Study
- ELLA: Generative AI-Powered Social Robots for Early Language Development at Home
- Vision-Language-Action (VLA) Models: Concepts, Progress, Applications and Challenges
Related papers
- Rethinking Actions: Continuity-Aware Priors for World-Model Planning
- MOSAIC: Modular Obstacle and Spatial-Aware Framework for Manipulation In Cluttered Scenes
- SAGAS: Semantic-Aware Graph-Assisted Stitching for Offline Temporal Logic Planning
- DexSeed: Expanding Human Hand-Object Interactions into Robot Demonstrations for Dexterous Policy Learning
- LaST0.5: A Generalist World Action Model with Interaction-Centric Latent Reasoning
- CIALLO: Learning Shared Intent Representations for Cross-Embodiment Manipulation