Skip to main content
Back to timeline
arXivSource publication:

More robot-policy samples stop paying off fast: PlanDiv splits candidates by latent plan, lifting CALVIN five-task chains from 56.9% to 69.9%

Synopsis

Test-time scaling for vision-language-action policies usually resamples one decoder and executes the best candidate, but those candidates differ only by noise: on CALVIN, going from four to eight adds just 0.06–0.07 completed tasks. PlanDiv changes where candidates come from: a lightweight planner proposes latent plan hypotheses, and a shared masked-diffusion executor decodes one action chunk per plan from a single backbone pass, with the policy executing the chunk its own likelihood ranks highest. With backbone, data and training recipe fixed, eight plan-conditioned candidates raise CALVIN ABCD five-task chain success from 56.9% to 69.9%, while autoregressive, plan-free and same-plan resampling reach at most 64.9%.

Source-provided article image: PlanDiv: Test-Time Scaling with Latent Plans for Vision-Language-Action Models
arXiv

Interpretation

Test-time scaling for robot policies no longer relies only on drawing more actions and scoring them; candidates now come from different latent plans, each standing for one way of handling the current situation. Earlier approaches resampled a single decoder, so candidates differed only by temperature noise, action perturbation or a rephrased instruction, and extra samples tended to inherit the same wrong direction. With backbone, data, training recipe and selector matched, three resampling sources — autoregressive, plan-free masked diffusion and a plan-conditioned head held to one fixed plan — gain only 0.06–0.07 completed tasks on CALVIN when going from four to eight candidates.

The change is an action head with a lightweight planner that maps learned seeds to latent plan hypotheses, plus a shared masked-diffusion executor that decodes one action chunk per plan from a single backbone pass. The masked-diffusion executor fills an action chunk by iterative unmasking rather than left-to-right generation; the plan only conditions the executor, and the backbone, tokenizer and action vocabulary stay unchanged. Planner and executor are trained jointly from demonstrations alone by latent responsibility assignment, a soft expectation-maximization-style step that credits each demonstration chunk to the hypotheses explaining it best, with no plan labels or rewards; on CALVIN ABCD eight plan-conditioned candidates reach 69.9%, against 64.4% for eight same-plan candidates from the same checkpoint.

The plan's contribution can be isolated: with decoder noise fixed and only the plan replaced, task outcomes change more than twice as often as when only the noise is replaced. Same-plan resampling shares the checkpoint, the first candidate and each candidate's random numbers with PlanDiv, so the gap between them can only come from the plans. On a bank of CALVIN sub-task start states, replacing only the plan changes the task outcome in 16.2% of states against 7.2% for replacing only the noise; eight different plans cover 92.1% of states against 86.1% for eight same-plan samples.

On a real Franka arm the gain appears on tasks with more than one solution, while a one-solution task is on par with a single-output baseline. The OpenVLA-OFT+ baseline outputs one action chunk per decision, whereas PlanDiv executes the highest-likelihood of eight candidates on the same backbone. The one-solution push task differs by one trial (25/25 against 24/25); on the two two-solution tasks PlanDiv succeeds in 51 of 54 trials against 40, removing 11 of the baseline's 14 failures.

Perspective

The demonstrated scope is two simulated benchmarks, CALVIN ABCD and LIBERO-Long, plus one Franka arm with two cameras running at 15 Hz; on hardware only three tasks were compared, two of which admit two solutions. Readers interested in other embodiments, task types or longer horizons should look for validation in those settings.

Plans are learned from demonstrations without plan labels or rewards, so the plan space is set by the data; on the one-solution task PlanDiv differs from the single-output baseline by a single trial, so the gain there is not pronounced. On the real robot, demonstrations for the two-solution tasks cover one side only partly: just 8 of 58 demonstrations for the can-around-obstacle task take the far side.