In Case You Missed It: The ICRA 2026 Recap

A team from the Dexterous AI group at Analog Devices, Inc. (ADI) hosted a booth and participated in the world’s leading robotic gathering, The IEEE International Conference on Robotics and Automation (ICRA). Held in Vienna, the event showcased the latest research advancements across mechanical, electronics, and AI research.  

ADI’s team, consisting of Tao Yu, Julie Russell, Wonju Lee, Matteo Grimaldi, Saleh Nabi, Wenjie Lu, and Jame Sun was on hand to present a demo of our novel tactile sensor prototype. The demo used an AI policy using a humanoid hand to trace a cable. The team trained the policy solely using tactile data, highlighting the sensor’s huge potential in advancing robotic dexterity.

A paper authored by Wonju Lee, Matteo Grimaldi, and Tao Yu was also accepted to the conference, proposing a Cross-Modal Transformer with physics-informed symmetry regularization for visuo-tactile fusion. It achieves 96.59% insertion success on the TacSL benchmark, nearly matching privileged force sensing while using only realistic tactile inputs.

Along with the demo and the paper, we attended many of ICRA’s sessions — and these were our favorite talks from the event.

Reframing Dexterity: From Policies to Physics and Metrics

MIT Professor and Eka Robotics co-founder Pulkit Agrawal’s  talk challenged a common assumption in robot learning: that better policies alone will unlock dexterity. Instead, he argued that real-world performance is governed by the full stack: learning, control, and physical embodiment. Agrawal’s point was that we are currently over-optimizing the first.

A central theme was the importance of throughput as the primary metric: not just whether a robot succeeds, but how quickly and reliably it can perform tasks. As highlighted in the slides, factors like spatial resolution or force sensitivity matter only insofar as they improve throughput, while higher-fidelity sensing (e.g., tactile) becomes critical for contact-rich, highly dexterous tasks.

Agrawal also emphasized that many current evaluation metrics are misleading. His group’s recent work on KaRMA (Kinematic Rolling Manipulation Ability) addresses this gap by measuring a hand’s ability to continuously reorient and translate objects within the grasp through rolling contact. This approach offers a much closer proxy to true dexterity than traditional metrics like DOF or workspace. This highlights a broader issue: the field often optimizes for what is easy to measure, rather than what actually matters for manipulation.

Finally, the talk reinforced that physics cannot be treated as a black box. Small changes in controller gains or contact dynamics can significantly affect outcomes, even with identical learned policies.

Key insight: Dexterity is not just a learning problem—it’s a systems problem, requiring co-design of hardware, control, and learning, along with better metrics to measure success.

Co-designing Robot Bodies and Intelligence

One of the more forward-looking talks came from University of Cambridge’s Rika Antonova, who presented the ADA‑GO (Adaptive Global Optimization) framework—part of the UK ARIA Robot Dexterity program. The core idea is to move beyond treating robot hardware and learning separately, and instead co-optimize fingertip design, materials, and manipulation policies together. As shown in the talk, ADA‑GO combines global optimization with reinforcement learning to search over both physical design parameters (e.g., fingertip geometry, compliance) and control strategies simultaneously, rather than relying on fixed hardware assumptions.

Early results were particularly compelling for task-specific dexterity. By automatically generating optimized soft fingertip designs, the system demonstrated robust grasping across challenging categories such as long thin objects, small flat items, and tiny objects—cases where traditional grippers often fail. This reinforces a broader theme seen across the workshop: dexterity is not just a policy problem, but an embodiment problem.

From a research perspective, ADA‑GO fits into a larger shift toward hardware–software co-design: instead of forcing learning algorithms to compensate for suboptimal hardware, the system searches for the best combined solution end-to-end. The project aims to scale this approach through hardware-in-the-loop optimization, where promising designs are evaluated in simulation and iteratively adapted to the real world—ultimately enabling faster discovery of low-cost, high-performance robot configurations.

Key insight: The next leap in manipulation may not come from better policies alone, but from jointly optimizing the robot’s “body” and “brain” as a single learning system.

Do We Still Need Dexterous Hands?

Stanford University’s Jeannette Bohg’s keynote tackled a provocative question: do we still need dexterous hands in the era of powerful learning-based systems? While simple two-finger grippers paired with large models can already achieve impressive results, she argued that the true value of dexterous hands lies not in mimicking humans, but in throughput and controllable subspace.

In practice, this means dexterous hands enable fast, in-hand manipulation and richer interaction strategies, such as stabilizing tools with one finger while manipulating them with others - capabilities that remain out of reach for simpler grippers. However, this flexibility comes at a cost: high-DOF hands are extremely difficult to control and even harder to collect teleoperation data for, due to frequent contact switching and complex dynamics. 

To address this, Bohg proposed a shift away from imitation of human hand motion toward object-centric learning. Her lab’s approach (Sim‑Tool‑Real / UniSim‑Real) trains a single RL policy to track object trajectories rather than hand poses, using pose sequences extracted from human videos. This allows one policy to generalize across tools and tasks, rather than requiring task-specific skill modules.

Building on this, the team’s Play‑to‑Effect framework uses a pretrained “play policy” of atomic skills, which can then be fine-tuned for precise tasks like assembly, improving robustness, learning speed, and recovery behavior.

Key insight: The future of dexterous manipulation may lie not in copying human hands, but in learning task-relevant object dynamics, interactions, and letting the controller figure out the rest.

Scaling Humanoid Intelligence: The “Data Sponge” Approach

Yuke Zhu’s talk focused on what he framed as the core bottleneck in humanoid robotics: data. The UT Austin professor and NVIDIA researcher thinks that while hardware and learning algorithms are rapidly improving, real-world robot data remains scarce and expensive, limiting progress beyond demos.

To address this, Zhu proposed a clear scaling strategy built on a concept he made famous – the data pyramid:

• Bottom layer: massive Internet-scale human video (cheap, abundant, passive)
• Middle layer: synthetic and simulated data (scalable, but imperfect)
• Top layer: real robot data (high quality, but limited)

The key idea is not to rely on any single source, but to combine all three, with a world model acting as a “data sponge”, absorbing and integrating heterogeneous data into a unified representation.
This aligns with the broader trend toward foundation models for robotics, where generalization comes from scale and diversity rather than task-specific datasets.

Zhu illustrated this with several systems. SONIC uses large-scale human motion capture data to train full-body humanoid controllers, greatly simplifying RL by using motion tracking instead of complex reward design. EgoScale learns manipulation from egocentric human videos via a three-stage pipeline (pretrain → align → fine-tune), achieving strong performance with less than 1% real robot data.

Perhaps most striking was the DreamZero world action model, which transforms video generation models into action generators, suggesting that AI-generated trajectories could approach the training value of real robot data.

Key insight: Scaling robot intelligence will require treating data as a system—combining human, synthetic, and real-world sources—while using world models to bridge the gaps between them.

We had a great time meeting ICRA attendees, sharing our thoughts, and collecting feedback about our demo. We look forward to participating in next year’s event to experience firsthand how the field as a whole progresses.