Vision–Language–Action Models / ICLR 2027 · Under review

PGOT: Physically Grounded
Optimal Transport

for Distilling Vision-Language-Action Models

Anonymous Authors · Under double-blind review

Physical structure. Compact policies. Consistent completion.7B / 9B teachers 0.8B students

01 / The idea

Action tokens have
physical meaning.

Action tokens encode physical commands, and individual decisions differ in their importance to task completion. PGOT combines physical distribution alignment with selective supervision: emphasize pivotal decisions and retain teacher guidance that agrees with the expert.

Abstract in brief

We introduce Physically Grounded Optimal Transport (PGOT), a distillation framework that grounds teacher–student alignment in physical action space and makes supervision selective.

Geometry measures action discrepancies through probability transport. Criticality emphasizes pivotal decisions in expert trajectories. Reliability attenuates teacher guidance that departs from demonstrated behavior.

Across CoNav-UAV and LIBERO-10, 0.8B students retain the success rates of their 7B/9B teachers with approximately 9–11× fewer parameters, while exceeding the teachers’ three-out-of-three completion.

Same probability mismatch.
Different physical distance.

Interactive
TeacherStudent
01Decoded physical action →
Forward KL1.367Unchanged
Physical W₁0.206Scales with distance

Illustrative 9-bin distributions on [0, 1], each with 0.60 peak probability and 0.05 in every other bin. PGOT uses 256 action tokens.

02 / The method

Three components.
One grounded objective.

All three components read decoded physical actions: the distance between commands, changes in expert behavior, and teacher–expert disagreement. PGOT requires only output distributions and offline demonstrations; deployment keeps the student’s architecture and inference unchanged.

G 01

Geometry

How far is the action mismatch?

Align action distributions on ordered dimensions with exact one-dimensional Wasserstein-1 distance. Moving probability farther in decoded physical action space incurs a larger cost; the binary gripper uses filtered FKL.

Physical action distance
C 02

Criticality

Where should learning be emphasized?

Use changes between neighboring expert commands to identify pivotal decisions. Bounded sample weights concentrate distillation on critical maneuvers and gripper transitions.

Action-transition weighting
R 03

Reliability

Which teacher guidance should transfer?

Filter teacher probability far from the expert action and reduce transport guidance when the teacher’s most likely action disagrees with the expert, limiting the transfer of distant teacher errors.

Expert-conditioned transfer
PGOT overview: UAV and robot action-token interfaces, a large teacher and compact student, and the Geometry, Criticality and Reliability training components.
Paper · Figure 1 A shared discrete-action VLA interface for UAV navigation and robotic manipulation. A frozen teacher transfers knowledge through PGOT to a compact student; Geometry, Criticality, and Reliability ground both distribution alignment and supervision in physical actions. Open figure ↗

03 / The evidence

Small students.
More consistent execution.

Using the same distillation coefficients in both domains, PGOT improves success and execution consistency over all evaluated distillation baselines. The larger gains are in three-out-of-three completion.

Download results

150 start–goal configurations · 3 executions each · 0–500 m · 200-step horizon · 0.5 m collision threshold

CoNav-UAV online evaluation. Bold marks the best student result in each column.
PolicySR (%) ↑Pass³ (%) ↑CR (%) ↓TR (%) ↓MMBC (m) ↑
Teacher SFT 7B87.56 ± 1.3978.679.782.675.00
Student SFT53.33 ± 1.1538.6744.002.672.78
FKL85.33 ± 1.3374.6711.782.895.27
RKL78.00 ± 2.3168.0016.006.005.00
ABKD80.89 ± 1.0267.3315.563.564.98
ToDi75.56 ± 3.4262.0017.786.675.00
Shallow-π85.33 ± 0.6774.0010.893.785.09
VLA-AD74.89 ± 0.3858.0020.894.224.20
PGOT Ours88.00 ± 2.9182.008.004.005.72
↗

+7.33 points in three-out-of-three completion; +2.67 points in success rate. Gains are over the strongest distillation baseline on each metric. PGOT also has the lowest collision rate and largest mean minimum building clearance among the evaluated students.

10 tasks × 50 initial states · 3 executions each · 520-step horizon

LIBERO-10 online evaluation. Bold marks the best student result in each column.
PolicySR (%) ↑Pass³ (%) ↑MES ↓
Teacher SFT 9B49.13 ± 0.6137.20408.02
Student SFT45.47 ± 0.9531.40409.95
FKL44.93 ± 1.3339.40413.11
RKL37.73 ± 0.8322.20433.07
ABKD41.93 ± 1.0327.40416.14
ToDi43.73 ± 0.8127.40416.21
Shallow-π47.07 ± 1.6232.40407.13
VLA-AD39.73 ± 0.9025.20424.94
PGOT Ours49.47 ± 0.6445.40400.47
↗

+6.00 points in three-out-of-three completion; +2.40 points in success rate. Gains are over the strongest distillation baseline on each metric. PGOT also uses the fewest mean execution steps, counting all episodes including failures.

SR: mean success rate ± sample standard deviation across three evaluation rounds. Pass³: fraction of initial configurations completed in all three executions. CR / TR: collision / timeout rate. MMBC: mean minimum building clearance. MES: mean execution steps over all episodes. Bold marks the best student value. Teacher SFT and Student SFT are supervised references. All students use Qwen3.5-0.8B; action tokens are sampled autoregressively over 256 bins at temperature 0.3. The three rounds repeat evaluation, not training. Source: paper, Table 1.

What does each component contribute?Component ablations

Removing any component lowers SR and Pass³ in both domains; in all six comparisons, the loss in Pass³ is larger. Geometry has the largest effect on both metrics. Without Reliability, LIBERO-10 loses 11.20 points of Pass³ but only 1.80 points of SR.

Component ablations, from paper Table 2.
ObjectiveUAV SR (%) ↑UAV Pass³ (%) ↑LIBERO SR (%) ↑LIBERO Pass³ (%) ↑
PGOT88.00 ± 2.9182.0049.47 ± 0.6445.40
w/o Geometry82.22 ± 2.1472.6744.60 ± 0.8729.80
w/o Criticality83.78 ± 1.5474.6748.67 ± 0.9042.40
w/o Reliability84.89 ± 3.2975.3347.67 ± 1.6234.20

Source: paper, Table 2. Removing Geometry sets w₁ = 0; Criticality sets γ = 0; Reliability sets α = βᵣ = 0.

Four matched UAV route comparisons at 141.69, 227.41, 364.62, and 408.75 meters, with three trials each for PGOT, FKL, and Shallow-pi. PGOT succeeds in all three trials on these selected tasks.
Matched UAV routes · Figure 2 Four tasks ordered by start–goal distance, with three trials per policy. Dark paths highlight a trial where PGOT succeeds and both baselines fail; light paths show the other two trials. These selected examples complement the full evaluation above. Open figure ↗

04 / In motion

From action tokens
to behavior.

Side-by-side replays compare PGOT, FKL, and Shallow-π on matched initial configurations. These are selected qualitative examples; the tables report the full benchmark evaluation.

01 / CoNav-UAV

UAV navigation

4 selected cases · 1080p / 60 fps
Jump to

Rows: PGOT, FKL, Shallow-π. Columns: first-person, third-person, and bird’s-eye views. In these four selected replays, PGOT succeeds while both baselines collide. Silent video on a shared policy-step timeline, normalized within each case; terminal frames are held.

Download video
02 / LIBERO-10

Robotic manipulation

3 selected cases · 1080p / 60 fps
Jump to

Columns: PGOT, FKL, Shallow-π. Rows: agent and wrist camera views. In these three selected replays, PGOT succeeds while both baselines time out. Silent video at 1× simulated time; terminal frames are held.

Download video

05 / A closer look

Understanding
the improvement.

Training and execution diagnostics show how physical transport corrects distant action errors, Criticality emphasizes difficult states, and Reliability limits the inheritance of unreliable teacher decisions.

Geometry-to-filtered-FKL gradient RMS ratios and gradient angles during training on CoNav-UAV and LIBERO-10.

Complementary transport gradients

Geometry and filtered FKL have mean gradient angles of 73.94° on CoNav-UAV and 72.05° on LIBERO-10. Transport gradient RMS is 15.13% and 9.94% of filtered FKL, supplying a distinct physical learning signal.

CoNav-UAV component diagnostics: tail probability mass and motion for Geometry; difficult states and late-turn outcomes for Criticality; teacher lean by action-bin distance and completion by teacher success count for Reliability.

Why the UAV student improves

Against w/o Geometry, PGOT reduces probability mass at least 32 bins from the expert by 17.26% under teacher forcing. Criticality emphasizes difficult maneuver states. On the 32 configurations where the teacher fails at least once in three trials, PGOT raises Pass³ from 43.75% to 62.50% over w/o Reliability.

Explore the research

Ground the objective.
Keep the student compact.

The anonymous repository includes training, data preparation, evaluation, and metrics.