Astra Sparks Debate: Has Scale Solved Computer Vision?

On September 20, a debate erupted on Twitter/X over whether scaled-up models have "solved" computer vision and other research fields. Proponents cited Google Astra's ability to perform detection and segmentation on arbitrary images in one pass, arguing that true intelligence is fundamentally a matter of scale and that the Bitter Lesson, having spread from NLP (around 2021) to vision, would next reach robotics within five years. The following day, discussion shifted to whether Astra's capabilities are genuinely new: Edgar Sucar acknowledged the long-standing foundations laid by the CV and robotics communities, noting his own prior work (e.g., NodeSLAM, 2020) sits in that lineage, and welcomed Astra building on shared research accumulation.

Confirmed

• Proponents (relayed by @Sadlylifes41u) argued that Astra's performance shows vision tasks can be done directly by scaled general models, and that intelligence is essentially a scale problem.
• CV researcher Georgia Chal countered that detection, segmentation and tracking are only part of vision; 3D/4D scene understanding, physical state and property estimation, contact and interaction understanding, uncertainty quantification, prediction under intervention, and robust generalization remain major open problems.
• Chal further noted that robotics especially exposes these shortcomings, making the "robotics solved in 5 years" claim unsupported.
• Responding to "AI has ended science," Chal argued that PhD-level tasks becoming benchmarks is not evidence of science ending but of a fast-moving frontier: general models push the frontier forward rather than invalidating research.
• Edgar Sucar acknowledged the field's foundational contributions, citing NodeSLAM (2020) as precedent showing Astra-style capabilities did not appear from nothing.

Unconfirmed

• "Robotics solved within five years" remains a one-sided prediction with no verifiable evidence from either side.
• Astra's actual capability boundaries (especially in 3D understanding, interaction, and prediction under intervention) lack systematic public evaluation; both sides relied on subjective experience or inference.

Why it matters

• The debate is a microcosm of the broader question of whether the Bitter Lesson applies to all fields: if vision and robotics are indeed "solved" by scaled general models, research funding, talent training, and technical roadmaps would need restructuring.
• The open problems listed by the opposing side—3D/4D understanding, physical estimation, interaction, causality, robustness—are precisely the core bottlenecks for robotics deployment, offering concrete reference points for judging the capability boundaries of general models.
• Sucar's precedent perspective reminds us that clear academic lineage exists between old and new systems; "breakthroughs" should be assessed within the context of accumulated research.

2026-09-20 ~ 2026-09-21 · 6 related posts

• [source] Is Scale 'Solving' Computer Vision? Researchers Clash Over 3D Understanding, Robotics Timeline — Sadlylifes41u · 2026-09-20
• [source] Is Scale 'Solving' Computer Vision? Researchers Clash Over 3D Understanding, Robotics Timeline — GeorgiaChal · 2026-09-20
• Vision researcher pushes back on 'field is solved' claim: robotics is not simply the next modality — GeorgiaChal · 2026-09-20
• Researchers clash: PhD problems becoming today's AI benchmarks shows progress, not an ending — GeorgiaChal · 2026-09-20
• [source] Researchers cite NodeSLAM (3DV 2020) as prior art for Astra-style robot skills — AjdDavison · 2026-09-21

1 near-duplicate retellings: Sadlylifes41u