At a glance
This viewpoint records a debate about whether large demonstration datasets and learned generative policies can deliver broadly capable robots. Its central disagreement concerns how much existing physical structure should guide learning, and whether empirical progress can precede a satisfactory theory.
The written positions converge more than the provocative title suggests: several researchers explicitly favor combinations of data and models. This is a collection of arguments and examples, not a controlled comparison establishing that either approach solves robotics.
Introduction: the debate and its scope
Nancy Amato and Seth Hutchinson organized the closing keynote debate at ICRA 2025; Ken Goldberg moderated it. The format assigned Garg, Rus and Tedrake to the affirmative side and Billard, Leslie Kaelbling and Park to the negative side, asking participants to take strong positions. The article contains five individual written arguments; it does not include a separate Kaelbling contribution.
More than 1,000 people attended. A show of hands suggested few changed their initial positions, but this was not a measured survey of the robotics community. The discussion concerned software and control, leaving hardware, materials, design and practical collection of massive datasets largely outside its scope [1] [2].

Animesh Garg: learning before complete theory
Garg argues that useful technology often arrives before a complete scientific explanation. General-purpose autonomy is underspecified: examples and feedback can define desirable behavior where no closed-form specification is available. Incomplete analytical models may also exclude good solutions from the outset.
He treats collection, curation and large-scale computation as scientific work, with software abstractions needed to make the resulting systems manageable. Foundation models might unify currently separate robotics problems through shared experience and physics. These are reasons to pursue data-driven methods; they do not establish a sufficient dataset size or demonstrate universal generalization.
Aude Billard: the interplay of models and data
Billard emphasizes that data acquire explanatory value through models. Her astronomy analogy highlights the productive cycle of carefully selected observations, hypotheses and revision rather than accumulation alone.
Transferring the breadth of human or animal physical knowledge to robots is harder than collecting language: available interfaces capture only part of embodied sensing and action, often with substantial burden. Even if brute-force computation delivers useful behavior, abandoning general principles could leave researchers with tool-specific recipes and limited understanding. Her conclusion favors an iterative relationship between gathering evidence and refining models.
Daniela Rus: combining physical structure and experience
Rus explicitly qualifies her assigned affirmative position: robust robotics needs both mathematical models and data. Physical structure is valuable in well-specified settings, while clutter, deformable objects, people and changing contexts introduce variation that simple analytical models do not fully capture.
She argues for rich embodied experience, including contact forces, joint torques, trajectories and gaze. Her laboratory’s kitchen testbed records body pose, muscle activity, forces and gaze during everyday tasks. The viewpoint describes this approach without reporting a controlled quantitative evaluation. Models provide constraints and physical consistency; diverse experience supplies coverage of real situations and recovery behavior.
Russ Tedrake: common sense for dexterous manipulation
Tedrake sees large-scale multitask pretraining as a promising route to physical common sense, including low-level recovery rather than only language interpretation. He cites a seminar and a separate empirical study reporting improved robustness, transfer with fewer demonstrations, and benefits from greater pretraining scale and diversity [3] [4].
A bimanual apple-preparation example illustrates his argument: demonstrations can teach subtle corrections when pieces slip, and pretraining may transfer recovery behavior from other tasks. The viewpoint supplies no new success-rate table or scaling experiment; those claims should be assessed in the cited study. He expects useful robots to precede a complete theory and treats eventual explanatory models as an ongoing objective.
Frank Park: robotics-specific inductive biases
Park cautions against assuming that language and vision scaling will transfer directly to robotics. Physical data are scarce, simulation is imperfect, embodiments vary, and deployment imposes demanding performance and resource constraints.
He proposes stronger inductive biases drawn from biology, physics, geometry and programming: control hierarchies, feedback, force and compliance models, useful symmetries and task-description languages [5]. Detailed manipulation data remain valuable, especially when they capture tactile feedback and compliant interaction. His objection is to scaling without appropriate structure, not to learning or foundation models themselves.
Convolutional networks and transformers also embody deliberate structural choices. Park argues that robotics needs comparably intentional designs rather than straightforward reuse of models built for other modalities.
Acknowledgments and original resources
The article preserves disagreement about priorities while showing broad support for combining experience with structure. It does not specify a common task distribution, reliability target or operational definition of “solving robotics,” so the positions are not directly testable alternatives under a shared protocol.
See the original ICRA debate video, the TRI multitask-transfer seminar, and the complete published PDF. Both videos remain on their original hosts. All five references and the full original figure are retained.