Sundae Robotics 12: What Scale Does Not Fix β Representation and Architecture in Robot Manipulation
βπ€π¨ Grab a sundae and join Sundae Robotics, a private, invite-only Sunday series bringing together robotics researchers, founders, and builders working at the frontier of physical intelligence.
βSundae Robotics 12
World Models, Contact-Rich Manipulation & Robot Foundation Models
Featured Talk: What Scale Does Not Fix β Representation and Architecture in Robot Manipulation
βWhat Scale Does Not Fix: Representation and Architecture in Robot Manipulation
βKeynote: Elijah Yi Herng Ong
Co-founder & CTO, Devol Robots
βRobot foundation models have scaled rapidly over the past two years, yet scale alone has not produced reliable physical interaction. In this talk, Elijah will present two architectures from Devol Robots that attack the problem from opposite ends: what a model represents, and how that model is organized.
βThe first is the Interaction World Model (IWM 1.0), which addresses representation. Built on Interaction Geometry, a Riemannian framework that keeps robot sensorimotor signals on the manifolds where they physically live rather than flattening them into Euclidean vectors, IWM infers compliance from the shape of a demonstration itself.
βThe model commands the robot through an impedance controller whose stiffness and damping it predicts, with no force or torque signals anywhere in training. On contact-rich tasks ranging from socket insertion to sub-millimeter lens seating, IWM reaches 92% average success, compared with 38% and 50% for leading VLA and world-model baselines, while using roughly an order of magnitude fewer action parameters.
βThe second is Devol-ONE, which addresses architecture. Existing world-action models keep prediction and policy structurally separate and connect them only through a predicted output. Devol-ONE instead runs vision-language reasoning, latent world prediction, and action generation as three streams of a single autoregressive Mixture of Transformers that exchange information at every layer.
βThis allows imagination to shape execution continuously rather than at a single bottleneck. Devol-ONE reaches 98.4% on LIBERO and 71.4% zero-shot on LIBERO-Plus without training on the perturbed data. On RoboTwin 2.0, it maintains performance under environment randomization while a strong pixel-space baseline falls from 76.6% to 2.7%.
βElijah will show how these two lines converge: a representation that captures the physics of contact, carried by an architecture that reasons about the future while it acts.
βTogether, they point toward what a deployable and reliable manipulation system can look like in industrial settings, and Elijah will demonstrate what each approach looks like running on production hardware today.
βElijah Yi Herng Ong is the Co-founder and CTO of Devol Robots, where he leads hardware, controls, and the embodied AI stack behind the company's force-control robotic assembly platform.
βAt Devol, Elijah works on robots designed for contact-rich industrial tasks such as plug insertion, hearing-aid assembly, snap-fits, alignment, and DIN-rail workβtasks where traditional position-controlled automation often struggles with uncertainty in force, geometry, and contact.
βHis research and engineering focus on enabling robots to understand physical contact and adapt in real time. His work spans embodied AI, reinforcement learning, force control, dexterous manipulation, grasping under uncertainty, and robot hardware-software co-design.
βElijah trained as a roboticist at Oregon State University, where his graduate research studied robotic grasping under uncertainty using deep reinforcement learning and human grasp demonstrations. His research examined how human-adapted grasp strategies could be transferred to robot hands and how geometric relationships between robot hands and objects could provide robust signals for manipulation.
βBefore founding Devol Robots, Elijah was a Senior Robotics Engineer at Roboligent, where he led work across robotics software, AI, learning-based perception, and low-level control for force-aware robotic manipulators.
βHis broader work reflects a recurring principle: robots operating in the physical world need to reason not only about where objects are, but about contact, compliance, alignment, uncertainty, and recovery as first-class parts of intelligence.
βTopics
ββ’ Why scaling robot foundation models does not automatically produce reliable physical interaction
ββ’ Representation and architecture as bottlenecks beyond model scale
ββ’ Interaction World Models for contact-rich robot manipulation
ββ’ Interaction Geometry and Riemannian representations of sensorimotor signals
ββ’ Preserving the physical structure of robot states and actions instead of flattening them into Euclidean vectors
ββ’ Inferring compliance directly from demonstrations
ββ’ Predicting impedance-controller stiffness and damping from learned representations
ββ’ Learning compliant behavior without force or torque supervision
ββ’ Contact-rich industrial tasks including insertion, alignment, and precision assembly
ββ’ Reducing action-space complexity while preserving dexterous behavior
ββ’ Devol-ONE and unified world-action architectures
ββ’ Joint vision-language reasoning, world prediction, and action generation
ββ’ Mixture of Transformers for embodied AI
ββ’ Continuous information exchange between reasoning, imagination, and control
ββ’ Latent world prediction versus pixel-space prediction
ββ’ Zero-shot robustness under visual and environmental perturbations
ββ’ LIBERO and LIBERO-Plus manipulation benchmarks
ββ’ RoboTwin 2.0 and robustness under environment randomization
ββ’ Bridging robot foundation models with impedance and force-aware control
ββ’ Deploying learned manipulation systems on production hardware
βOpen Discussion + Q&A
ββ’ Why has scaling vision-language-action models not translated directly into reliable contact-rich manipulation?
ββ’ What properties of physical interaction are lost when robot states and actions are represented as ordinary Euclidean vectors?
ββ’ When is a Riemannian representation materially better than a conventional learned latent space?
ββ’ Can compliance really be inferred from the geometry of a demonstration without explicit force or torque measurements?
ββ’ What does a manipulation model need to understand about contact that vision alone cannot provide?
ββ’ Should robot foundation models output joint positions, trajectories, forces, impedances, or some combination of them?
ββ’ How much of manipulation reliability comes from the learned model versus the underlying controller?
ββ’ Why might predicting stiffness and damping produce better contact behavior than directly predicting actions?
ββ’ Are current robot action representations unnecessarily high-dimensional?
ββ’ Should world prediction and action generation remain separate modules, or should they be learned inside a single architecture?
ββ’ What is lost when a world model communicates with a policy through only one predicted output?
ββ’ Can reasoning, world prediction, and control meaningfully improve one another when they exchange information at every transformer layer?
ββ’ Is latent-space prediction inherently more robust than pixel-space prediction for robot manipulation?
ββ’ What kinds of perturbations reveal whether a manipulation policy has learned genuine physical structure rather than visual correlations?
ββ’ How should manipulation benchmarks evolve to measure reliability under real-world variation rather than only nominal task success?
ββ’ What changes when the goal shifts from impressive robot demonstrations to systems that must run reliably on production hardware?
ββ’ Can one architecture eventually combine semantic reasoning, physical world modeling, compliance, and low-level control into a single deployable system?
ββ’ Beyond more data and larger models, what representations and architectural inductive biases will matter most for the next generation of robot foundation models?
