Cambridge startup Eka uses sim-based reinforcement learning and custom touch sensors to build a robotic claw that handles arbitrary objects with near-human dexterity.
Key Takeaways
Eka’s cofounders Pulkit Agrawal (MIT) and Tuomas Haarnoja (ex-Google DeepMind) train robots entirely in simulation, no human demonstration data, closer to AlphaZero than VLA models.
Their novel vision-force-action model embeds physics (mass, inertia) and tactile feedback, not just visual input, to close the sim-to-real gap.
Demos include screwing in a light bulb, grabbing arbitrary objects (keys, hairbrushes, earplugs), and packing chicken nuggets on a moving conveyor with improvised tossing when needed.
Food handling and warehouse picking are the near-term commercial targets; founders claim superhuman dexterity, not just human-level, as the goal.
Authors compare current capability to GPT-1: glimmers of general physical intelligence, not yet proven at scale.
Hacker News Comment Review
Skeptics note that sim-only training for dexterous picking is not novel; many funded startups have pursued it for years, making Eka’s claimed edge hard to evaluate without disclosed technical details.
The real validation bar cited is Amazon bin-picking at production scale, which has resisted years of well-funded attempts and remains unsolved.
Safety concern raised: industrial robot arms that can fall or apply force near children or untrained workers are a harder adoption barrier than capability alone.
Notable Comments
@Animats: Points to Amazon’s long-running bin-picking failure as the only credible production benchmark for claims like Eka’s.
@NalNezumi: Argues sim-based picking training is already crowded and Eka’s differentiation is unclear without disclosed methods.
