The gap between a policy that works in a lab and a policy that works in a warehouse is not model scale. It is not data volume. It is the ability to take a long-horizon task, decompose it under uncertainty, and execute the pieces at control frequency without breaking anything when the world moves — which it always does.
I want to spend this internship working on that gap, inside the π-series stack.
What I bring is not a claim that I know the answer. It is evidence that I have been asking the right questions, at a scale I could afford, with results that survive scrutiny.
The data side, first
My most recent published work is a controlled study of the robot data flywheel — the closed loop in which a deployed policy's rollouts are collected, curated, and fed back into retraining. Every serious scaling bet in this field — RT-2, Open X-Embodiment, DROID, and the deployment programs that followed — assumes that loop compounds. I built the smallest faithful implementation and varied only the curation strategy. The result: with no feedback, a frozen policy sits at 0.12 held-out success and never moves. Self-curation (keeping your own successes) plateaus — it densifies what you already do. Relabeling deployment failures with an oracle compounds: 0.15 → 0.66, four times the success, at a fraction of the interaction budget of RL from scratch. Curated relabeling bought 1.4× the performance per label, and it was robust to label noise in a way blind relabeling was not. Every number in that paper regenerates from a committed JSON. I did not claim the result; I committed it.
That finding is a thesis about why scaling works in this field, and it has a direct consequence for π0-class models: the marginal value of an episode depends on the curation rule, and the highest-leverage investment in the π stack is not only more compute — it is a scoring layer that decides which deployment failures deserve an expensive label, and an oracle channel that can produce it. I would like to spend part of this internship testing that at your scale, where rollout telemetry is orders of magnitude richer than anything I could generate.
The policy side
I have trained and deployed diffusion policies and VLA architectures end-to-end: data collection and teleoperation pipelines, action chunking with receding-horizon execution, sim-to-real with domain randomization, and the less glamorous engineering — calibration, gripper failure modes, the intervention rate as a number you actually track rather than a hope. On real hardware I have learned the metric that papers underreport: not how often a policy succeeds, but how often a human has to stop it. That number is the operating cost of an embodied system, and it is the number I design against.
Why Physical Intelligence, and why the π-series
Because π0 got two architectural decisions right that most of the field has still not internalized. First, flow matching over diffusion for action generation: few-step sampling that keeps trajectories smooth at control frequency, in a formulation that is an ODE — a velocity field you can analyze, steer, and certify, not a stochastic sampler you can only sample from. Second, the VLM as the hierarchy: language and visual subgoal tokens flowing through the same backbone that produces actions, so semantic reasoning and motor control share one learned representation instead of two models awkwardly glued together. π0.7 scaling that across embodiments and object types is, to my reading, the strongest public evidence that this parameterization is the right substrate for generalist manipulation.
The proposal attached to this application is what I would do with an internship: add the third decision. A Steerable Multimodal Conditioning layer — a gated adapter that lets human corrections, world-model replans, and constraint setpoints steer the flow field mid-execution, with a Grönwall-type bound certifying that steering cannot jerk the robot — wrapped in a control-barrier-function safety envelope that admits steerability only through a formally verified filter. I chose the three benchmark tasks (cable untangling, textile folding on shifting fabric, adaptive tool use in clutter) to be the ones that defeat both hardcoded logic and monolithic imitation: non-rigid, high-entropy, and long-horizon, exactly where subgoal factorization and steerability stop being conveniences and become necessary.
What I am not asking for: a project to be handed to me, or credit for ideas someone else already had. I am asking for access to the hardest version of a problem I have already started solving, and for the chance to be wrong in public, quickly, in a place where being wrong is cheap and the data loop is real.
I build things that compile, measure things I can defend, and write papers where every number traces to a committed artifact. Give me a bench, a teleop rig, and the π stack, and I will give you back a steerable, verifiable action expert and the ablation study that says whether it was worth it.