Thanks! That would definitely be an interesting experiment.
One small clarification, though. we don’t compare against π₀ in this paper, and for the standard OpenVLA, as far as I understood LIBERO results use separately fine-tuned checkpoints for each suite. Our setup follows the same suite-specific training protocol.
That said, I’m also very curious about the joint-training setting too though. Definitely something worth testing!
That’s a good point. For the real-robot extension, we can separate task-level novelty from pretraining-level novelty. Held-out everyday objects would only test novelty relative to the robot demonstrations, since OpenVLA may already know those categories from pretraining.
One way around that is to include custom physical objects or scenes created after pretraining, which at least guarantees that the exact instances could not have appeared in its corpus.
That’s exactly the boundary we’re interested in. Our controlled baselines suggest that the result is not simply because LIBERO can be solved by any 0.68M-parameter policy: similarly sized Transformers and LSTMs perform much worse under the same front end, demonstrations, and decoder.
So our current claim is narrower, predictive sensorimotor structure can dramatically reduce the capacity needed for these tasks. Where large pretrained VLAs should plausibly pull ahead is under semantic/visual distribution shift, such as genuinely novel objects, scenes, or instructions. That’s an important next comparison.
[ New Predictive VLA policy published! ]
I would like to share our recent study just published several days ago on arXiv which questions: Do robot policies really need Transformers or Diffusion?
We introduce PredVLA: a 0.68M-parameter predictive-coding RNN with no robot-data pretraining, achieving comparable success rate to OpenVLA.
Instead of learning observation → action, PredVLA predicts sensorimotor dynamics and uses prediction errors to drive action inference.
Maybe compact robot policies don’t need more parameters.
Maybe they need a better learning principle.
📄 https://t.co/nh4iYWIyyo
#Robotics #EmbodiedAI #VLA #PredictiveCoding #RobotLearning
Introducing GEN-1.
Our latest milestone in scaling robot learning.
We believe it to be the first general-purpose AI model to master simple physical tasks.
99% success rates, 3x faster speeds, adapts in real time to unexpected scenarios, w/ only 1 hour of robot data.
More🧵👇