Robotics Researchers Debate: Task-Specific Hacks Are Dead, Attack Scaling Instead

ChongZzZhang · x · 2026-11-22

An opinionated exchange between ChongZzZhang and JohannKing4 on robotics research direction.

JohannKing4 asks: if strong limited priors hacked onto a small task set make a robot work well, is that really hacking? Task-focused research used to dominate robotics, letting grads understand real bottlenecks instead of chasing another benchmark.

ChongZzZhang counters that pre-LLM that thesis made sense, but now almost no single task needs something new — stacking existing engineering know-how suffices, and agents make it trivial. JohannKing4's summary of viable directions: drop A+B-style compositions and prior-hacking; solve real bottlenecks (scaling and performance upper bounds); or make agents/VLMs significantly better.

Related event: Researchers Debate Whether Single-Task Hacks Are Obsolete in the LLM Era(2 posts)→

Original post →

More from Embodied

Embodied channel →