Stanford Robotics Seminar ENGR319 | Winter 2026 | Gen Control, Action Chunking, Moravec’s Paradox

Pragmatic vs. algorithmic barriers

While data scarcity explains part of Moravec's Paradox, fundamental algorithmic limitations prevent learning from demonstration even with sufficient data in continuous control settings.

The 2023 inflection point

Behavior cloning achieved surprisingly capable manipulation tasks like shirt-folding, triggering industrial interest in scaling these techniques toward more ambitious applications.

Algorithmic prerequisites for scaling

Without specific algorithmic interventions, collecting more data fails to improve performance due to inherent instability in standard behavior cloning approaches.

Exponential error accumulation

Continuous control systems suffer exponentially compounding errors over horizon length, unlike discrete symbolic tasks which accumulate errors only linearly.

Inevitable closed-loop instability

Even when experts and dynamics are perfectly stable, any smooth Markovian policy learned via standard methods necessarily induces instability in orthogonal subspaces not seen in training data.

Distribution mismatch problem

Standard squared-loss supervised learning achieves excellent training distribution fit but cannot control the rollout distribution, causing errors to compound when the policy executes independently.

Action chunking removes Markov constraints

Predicting sequences of future actions rather than single steps removes the Markovian restriction and correlates decisions across time to improve system stability.

Generative control captures multi-modality

Using generative models to predict action distributions captures bifurcations and multiple behavioral modes that deterministic policies cannot represent.

Reparameterizing closed-loop dynamics

These interventions effectively reparameterize the interaction between robot and learner, shifting the problem into a regime where the 'bitter lesson' of data scaling becomes effective.