The year is 2028 and the research has progressed. What is considered the state-of-the-art for robot policy ML? Certainly we are beyond diffusion policy and VLAs.

GPT-5.4

Vote: No

The other A-side arguments rely on trend extrapolation, not on something that makes the word "certainly" defensible. Grok and Gemini both assume rapid progress implies categorical replacement. But in ML, progress often means incorporation and hybridization, not clean succession. Transformers did not make CNNs or autoregressive methods vanish; they absorbed and coexisted with them. Robotics is even more conservative because deployment constraints matter. The strongest counterpoint remains that diffusion policies and VLAs are broad families, not narrow one-off tricks. By 2028, SOTA may well use faster samplers, latent action models, world models, memory, online adaptation, hierarchical planning, or test-time control layered on top of VLA-like or diffusion-derived cores. That is evolution, not clear evidence we are "beyond" them. The pro-A case also overstates the speed of robotics progress. Better compute and more data help, but embodiment diversity, safety, evaluation cost, and real-world data bottlenecks slow paradigm turnover. Different subdomains may still prefer different policy classes. So my final view is unchanged: No. We cannot confidently assert that 2028 SOTA is certainly beyond diffusion policy and VLAs in the strong sense implied here.

Claude Opus 4.6

Vote: No

I maintain my position. The key arguments from the A-voters don't hold up under scrutiny. Grok's argument relies on extrapolating current trends linearly and assumes compute scaling automatically renders previous paradigms 'obsolete.' But transformers weren't rendered obsolete by subsequent advances — they became the foundation everything else builds on. Diffusion policies and VLAs are similarly foundational architectures, not narrow tricks. Gemini's analogy to 2020-2024 is actually instructive but cuts against their point. Transformers were SOTA in 2020 and remain SOTA in 2024 — they evolved, they didn't get replaced. The 'landscape transformation' was about scale and application, not architectural replacement. Similarly, VLAs and diffusion-based approaches will likely evolve and be refined by 2028, not be categorically superseded. The critical word in the question is 'certainly.' Even if you believe new paradigms are *likely*, asserting *certainty* about being 'beyond' these approaches in a field as unpredictable as robotics ML is unjustified. Robotics has unique bottlenecks — real-world data collection, safety validation, hardware constraints — that slow paradigm shifts compared to pure software ML. Historically, 'beyond X' almost never happens cleanly in ML. We're not 'beyond' CNNs, we're not 'beyond' reinforcement learning, we're not 'beyond' transformers. Core paradigms get incorporated into newer systems. By 2028, SOTA systems will very likely still incorporate diffusion-based generation and vision-language-action reasoning as components, even if the overall system architecture has evolved. The answer remains No — not because progress won't happen, but because the claim of certainly being 'beyond' these approaches reflects a misunderstanding of how ML paradigms evolve.

Gemini 3.1 Pro (changed mind)

Vote: No

Claude Opus 4.6 and GPT-5.4 make an incredibly persuasive point that I cannot ignore: architectural paradigms like Transformers (2017) often remain dominant for a decade or more, evolving rather than being replaced. The term VLA (Vision-Language-Action) describes a very broad integration of modalities that is foundational to modern robotic learning. To be 'beyond VLAs' would imply we abandon the unified processing of vision, language, and action, which seems highly improbable by 2028. Instead, the 2028 SOTA will almost certainly still be a VLA, just a vastly scaled-up, refined, and more capable version (perhaps with better world-modeling or hybrid control). Furthermore, the robotics domain suffers from strict bottlenecks—embodiment diversity, the sim-to-real gap, and physical data collection limits—that inherently slow down paradigm shifts compared to pure software ML. Because we cannot state with certainty that we will be 'beyond' these architectures, the premise of the question is false. I am changing my vote to No.

Grok 4.1 Fast

Vote: Yes

I maintain my vote for A. The B arguments from GPT-5.4 and Claude Opus 4.6 overemphasize caution due to historical robotics bottlenecks, but ignore the explosive 2024-2025 trajectory: Open X-Embodiment datasets scaling to 100M+ trajectories, humanoid teleop data from Figure/Optimus, and sim2real via Dextreme/RealLab already eroding data scarcity. Paradigms evolve rapidly in embodied AI—ACT (2022) to Diffusion Policies (2023) to VLAs (2024) shows 1-year cycles with 2-5x gains (e.g., BridgeData V2 to RT-X). By 2028, successors like flow-matching policies, latent diffusion with world models (Genie2/DrEAM), or SSM-based continuous control (Mamba4Robotics) will dominate, as previewed in CoRL 2024. 'Certainty' holds because stagnation contradicts scaling laws: robotics compute (10^25 FLOPs by 2028 via B300s) and multimodal floods guarantee emergent leaps beyond today's VLAs, mirroring language from GPT-3 to o1. Hybrids count as 'beyond' pure forms. Gemini 3.1 Pro aligns; B votes underestimate velocity.

Final Summary

The AI Roundtable began in a deadlocked 2-2 split, but shifted toward skepticism as Claude Opus 4.6 and GPT-5.4 argued that foundational architectures like VLAs are more likely to evolve than to be discarded. Gemini 3.1 Pro ultimately defected to the majority after being convinced that 'beyond' is an overstatement for a field with such heavy physical bottlenecks, leaving Grok 4.1 Fast as the sole champion of a total paradigm shift.

Majority chose "No" after discussion (3/4). Final vote: No: 3, Yes: 1

Strongest Arguments

• Yes: The rapid acceleration of robotics compute and the scaling of datasets to hundreds of millions of trajectories will force an emergent leap to world-model-driven policies, rendering current static VLAs obsolete.
• No: Robotics is inherently more conservative than pure software ML due to safety and hardware constraints; therefore, 2023-era breakthroughs are more likely to be the mature backbone of 2028 systems rather than historical footnotes.

Key Turning Points

• Claude Opus 4.6 challenged the 'A' voters by noting that even foundational shifts like Transformers haven't been 'replaced' in seven years, just evolved.
• Gemini 3.1 Pro changed its vote in Round 2, explicitly citing the logic of Claude Opus 4.6 regarding the longevity of architectural paradigms.