The RLDX-1 discussion reinforces a critical shift in robotics: embodied intelligence will not be solved by vision-language scale alone, but by integrating tactile, force, temporal, and high-DoF interaction data. Dexterity is not downstream of intelligence; it is one of the core pathways through which physical AI becomes commercially reliable.
Agreed! How can we have world models that don't incorporate more modalities of sensing! This goes all the way back to Craik's theory that cognition consists of mental models of the world allowing us to experiment internally. (ht @hugo @robonaissance for reminding me)
The line that scale cannot recover a modality it never received applies beyond robotics. In knowledge work, a model may see the document but miss the conversation, exception, or field judgment that gave it meaning. Benchmarking should therefore test whether the system recognizes absent evidence, not only whether it succeeds on complete inputs. Do current dexterity benchmarks reward an explicit 'insufficient sensing' response, or do they still push models to act even when the decisive modality is unavailable?
The RLDX-1 discussion reinforces a critical shift in robotics: embodied intelligence will not be solved by vision-language scale alone, but by integrating tactile, force, temporal, and high-DoF interaction data. Dexterity is not downstream of intelligence; it is one of the core pathways through which physical AI becomes commercially reliable.
Agreed! How can we have world models that don't incorporate more modalities of sensing! This goes all the way back to Craik's theory that cognition consists of mental models of the world allowing us to experiment internally. (ht @hugo @robonaissance for reminding me)
The line that scale cannot recover a modality it never received applies beyond robotics. In knowledge work, a model may see the document but miss the conversation, exception, or field judgment that gave it meaning. Benchmarking should therefore test whether the system recognizes absent evidence, not only whether it succeeds on complete inputs. Do current dexterity benchmarks reward an explicit 'insufficient sensing' response, or do they still push models to act even when the decisive modality is unavailable?
Absolutely true. Benchmarking awareness of absence is a great idea.