Skip to content
ixian
Optimize this model
← Model library

v0.5-libero

VLA · Verne Robotics · measured on AGX Thor

Benchmark results for the model from vernerobotics/v0.5-libero.

Model repository ↗

Compare with other models on AGX Thor →

Performance highlights

PyTorch eager → ZETIC optimized · AGX Thor

Policy latency · p50
79.2%4.8× faster
Energy per inference
73.3%Less energy
Peak memory
27.2%Less memory

Performance comparison

Lower is better

AGX Thor · per policy call

Recorded measurements for the selected reference and ZETIC optimized paths.
MetricReferencePyTorch eager · BF16 · 4 flow stepsZETIC optimizedMixed precision · 2 flow steps
Policy latency · p50259.39 ms53.92 ms
Energy per inference22.14 J5.9 J
Peak memory44.02 GB32.04 GB

Different inference configurations.

View optimized build results →

Measurement conditions and source

Peak board memory above idle (MemTotal − MemAvailable), over the full command including model loading; not process RSS.

p50. v0.5-libero · libero · AGX Thor

Speedup is reference latency divided by path latency on this board. Lowest recorded latency excludes marked model or inference-schedule variants. Check the configurations below before treating paths as a matched comparison.

PyTorch eager · Model authors · BF16

PyTorch eager · BF16 · 4 flow steps · batch 1

Peak board memory above idle (MemTotal − MemAvailable), over the full command including model loading; not process RSS.

ZETIC · ZETIC · Mixed precision

ZETIC · Mixed precision · 2 flow steps · batch 1 · Different configuration

Additional validation pending. Peak board memory above idle (MemTotal − MemAvailable), over the full command including model loading; not process RSS.

PyTorch eager: End-to-end latency, p50 (median of round medians), 4 rounds · 232 samples per round.

ZETIC: End-to-end latency, p50 (median of round medians), 4 rounds · 1,076 samples per round.

Source: reviewed benchmark measurements. Public measurement: record-572e80f79750962162d0c6af18974e62396c80819087d46bc4bfb2ef1912ac38.

Measurement methodology ↗

Quality evaluation

v0.5-libero · LIBERO Goal

AGX Thor · Same episodes for both paths

PyTorch eager

BF16

99%198/200

Backend
PyTorch eager
Precision
BF16
Tasks
10
Episodes
200
Flow steps
4

ZETIC optimized

MIXED PRECISION

99%198/200

Backend
ZETIC
Precision
MIXED PRECISION
Tasks
10
Episodes
200
Flow steps
2
Checkpoint
vernerobotics/v0.5-libero
Protocol
Closed loop
Seeds
PyTorch eager · Seed 7: 198/200ZETIC optimized · Seed 7: 198/200
Source
Per-task outcomes are listed below.

Observed difference: 0 pp

Evaluation details and episode results
Per-task success and episode outcomes: O = success, x = failure.
Path / seedTask / taskResultFailed episodes
PyTorch eagerSeed 70 · Task 020/20OOOOOOOOOOOOOOOOOOOONone
PyTorch eagerSeed 71 · Task 120/20OOOOOOOOOOOOOOOOOOOONone
PyTorch eagerSeed 72 · Task 220/20OOOOOOOOOOOOOOOOOOOONone
PyTorch eagerSeed 73 · Task 320/20OOOOOOOOOOOOOOOOOOOONone
PyTorch eagerSeed 74 · Task 420/20OOOOOOOOOOOOOOOOOOOONone
PyTorch eagerSeed 75 · Task 519/20OOOOOOOOOOOOOOOOOOOx19
PyTorch eagerSeed 76 · Task 620/20OOOOOOOOOOOOOOOOOOOONone
PyTorch eagerSeed 77 · Task 720/20OOOOOOOOOOOOOOOOOOOONone
PyTorch eagerSeed 78 · Task 820/20OOOOOOOOOOOOOOOOOOOONone
PyTorch eagerSeed 79 · Task 919/20OOOOxOOOOOOOOOOOOOOO4
ZETIC optimizedSeed 70 · Task 020/20OOOOOOOOOOOOOOOOOOOONone
ZETIC optimizedSeed 71 · Task 120/20OOOOOOOOOOOOOOOOOOOONone
ZETIC optimizedSeed 72 · Task 220/20OOOOOOOOOOOOOOOOOOOONone
ZETIC optimizedSeed 73 · Task 320/20OOOOOOOOOOOOOOOOOOOONone
ZETIC optimizedSeed 74 · Task 420/20OOOOOOOOOOOOOOOOOOOONone
ZETIC optimizedSeed 75 · Task 520/20OOOOOOOOOOOOOOOOOOOONone
ZETIC optimizedSeed 76 · Task 620/20OOOOOOOOOOOOOOOOOOOONone
ZETIC optimizedSeed 77 · Task 720/20OOOOOOOOOOOOOOOOOOOONone
ZETIC optimizedSeed 78 · Task 820/20OOOOOOOOOOOOOOOOOOOONone
ZETIC optimizedSeed 79 · Task 918/20OOOOxOOOOOOOOOOOxOOO4, 16

O = success · x = failure.

v0.5-libero · LIBERO Spatial

AGX Thor · Same episodes for both paths

PyTorch eager

BF16

98.5%197/200

Backend
PyTorch eager
Precision
BF16
Tasks
10
Episodes
200
Flow steps
4

ZETIC optimized

MIXED PRECISION

98.5%197/200

Backend
ZETIC
Precision
MIXED PRECISION
Tasks
10
Episodes
200
Flow steps
2
Checkpoint
vernerobotics/v0.5-libero
Protocol
Closed loop
Seeds
PyTorch eager · Seed 7: 197/200ZETIC optimized · Seed 7: 197/200
Source
Per-task outcomes are listed below.

Observed difference: 0 pp

Evaluation details and episode results
Per-task success and episode outcomes: O = success, x = failure.
Path / seedTask / taskResultFailed episodes
PyTorch eagerSeed 70 · Task 020/20OOOOOOOOOOOOOOOOOOOONone
PyTorch eagerSeed 71 · Task 118/20OOOOxOOOOOOOOOxOOOOO4, 14
PyTorch eagerSeed 72 · Task 220/20OOOOOOOOOOOOOOOOOOOONone
PyTorch eagerSeed 73 · Task 320/20OOOOOOOOOOOOOOOOOOOONone
PyTorch eagerSeed 74 · Task 419/20OOOOOOOOOOOOOOOOOOxO18
PyTorch eagerSeed 75 · Task 520/20OOOOOOOOOOOOOOOOOOOONone
PyTorch eagerSeed 76 · Task 620/20OOOOOOOOOOOOOOOOOOOONone
PyTorch eagerSeed 77 · Task 720/20OOOOOOOOOOOOOOOOOOOONone
PyTorch eagerSeed 78 · Task 820/20OOOOOOOOOOOOOOOOOOOONone
PyTorch eagerSeed 79 · Task 920/20OOOOOOOOOOOOOOOOOOOONone
ZETIC optimizedSeed 70 · Task 020/20OOOOOOOOOOOOOOOOOOOONone
ZETIC optimizedSeed 71 · Task 118/20OOOOxOOOOOOOOOxOOOOO4, 14
ZETIC optimizedSeed 72 · Task 220/20OOOOOOOOOOOOOOOOOOOONone
ZETIC optimizedSeed 73 · Task 320/20OOOOOOOOOOOOOOOOOOOONone
ZETIC optimizedSeed 74 · Task 419/20OOOOOxOOOOOOOOOOOOOO5
ZETIC optimizedSeed 75 · Task 520/20OOOOOOOOOOOOOOOOOOOONone
ZETIC optimizedSeed 76 · Task 620/20OOOOOOOOOOOOOOOOOOOONone
ZETIC optimizedSeed 77 · Task 720/20OOOOOOOOOOOOOOOOOOOONone
ZETIC optimizedSeed 78 · Task 820/20OOOOOOOOOOOOOOOOOOOONone
ZETIC optimizedSeed 79 · Task 920/20OOOOOOOOOOOOOOOOOOOONone

O = success · x = failure.

v0.5-libero · LIBERO Object

AGX Thor · Same episodes for both paths

PyTorch eager

BF16

99.5%199/200

Backend
PyTorch eager
Precision
BF16
Tasks
10
Episodes
200
Flow steps
4

ZETIC optimized

MIXED PRECISION

99.5%199/200

Backend
ZETIC
Precision
MIXED PRECISION
Tasks
10
Episodes
200
Flow steps
2
Checkpoint
vernerobotics/v0.5-libero
Protocol
Closed loop
Seeds
PyTorch eager · Seed 7: 199/200ZETIC optimized · Seed 7: 199/200
Source
Per-task outcomes are listed below.

Observed difference: 0 pp

Evaluation details and episode results
Per-task success and episode outcomes: O = success, x = failure.
Path / seedTask / taskResultFailed episodes
PyTorch eagerSeed 70 · Task 020/20OOOOOOOOOOOOOOOOOOOONone
PyTorch eagerSeed 71 · Task 120/20OOOOOOOOOOOOOOOOOOOONone
PyTorch eagerSeed 72 · Task 220/20OOOOOOOOOOOOOOOOOOOONone
PyTorch eagerSeed 73 · Task 320/20OOOOOOOOOOOOOOOOOOOONone
PyTorch eagerSeed 74 · Task 420/20OOOOOOOOOOOOOOOOOOOONone
PyTorch eagerSeed 75 · Task 519/20OxOOOOOOOOOOOOOOOOOO1
PyTorch eagerSeed 76 · Task 620/20OOOOOOOOOOOOOOOOOOOONone
PyTorch eagerSeed 77 · Task 720/20OOOOOOOOOOOOOOOOOOOONone
PyTorch eagerSeed 78 · Task 820/20OOOOOOOOOOOOOOOOOOOONone
PyTorch eagerSeed 79 · Task 920/20OOOOOOOOOOOOOOOOOOOONone
ZETIC optimizedSeed 70 · Task 020/20OOOOOOOOOOOOOOOOOOOONone
ZETIC optimizedSeed 71 · Task 120/20OOOOOOOOOOOOOOOOOOOONone
ZETIC optimizedSeed 72 · Task 220/20OOOOOOOOOOOOOOOOOOOONone
ZETIC optimizedSeed 73 · Task 320/20OOOOOOOOOOOOOOOOOOOONone
ZETIC optimizedSeed 74 · Task 420/20OOOOOOOOOOOOOOOOOOOONone
ZETIC optimizedSeed 75 · Task 519/20OOOOOOxOOOOOOOOOOOOO6
ZETIC optimizedSeed 76 · Task 620/20OOOOOOOOOOOOOOOOOOOONone
ZETIC optimizedSeed 77 · Task 720/20OOOOOOOOOOOOOOOOOOOONone
ZETIC optimizedSeed 78 · Task 820/20OOOOOOOOOOOOOOOOOOOONone
ZETIC optimizedSeed 79 · Task 920/20OOOOOOOOOOOOOOOOOOOONone

O = success · x = failure.

v0.5-libero · LIBERO 10

AGX Thor · Same episodes for both paths

PyTorch eager

BF16

98.5%197/200

Backend
PyTorch eager
Precision
BF16
Tasks
10
Episodes
200
Flow steps
4

ZETIC optimized

MIXED PRECISION

99%198/200

Backend
ZETIC
Precision
MIXED PRECISION
Tasks
10
Episodes
200
Flow steps
2
Checkpoint
vernerobotics/v0.5-libero
Protocol
Closed loop
Seeds
PyTorch eager · Seed 7: 197/200ZETIC optimized · Seed 7: 198/200
Source
Per-task outcomes are listed below.

Observed difference: +0.5 pp

Evaluation details and episode results
Per-task success and episode outcomes: O = success, x = failure.
Path / seedTask / taskResultFailed episodes
PyTorch eagerSeed 70 · Task 020/20OOOOOOOOOOOOOOOOOOOONone
PyTorch eagerSeed 71 · Task 120/20OOOOOOOOOOOOOOOOOOOONone
PyTorch eagerSeed 72 · Task 220/20OOOOOOOOOOOOOOOOOOOONone
PyTorch eagerSeed 73 · Task 319/20OOOxOOOOOOOOOOOOOOOO3
PyTorch eagerSeed 74 · Task 420/20OOOOOOOOOOOOOOOOOOOONone
PyTorch eagerSeed 75 · Task 519/20OOOOOOOOOOOOOOxOOOOO14
PyTorch eagerSeed 76 · Task 620/20OOOOOOOOOOOOOOOOOOOONone
PyTorch eagerSeed 77 · Task 720/20OOOOOOOOOOOOOOOOOOOONone
PyTorch eagerSeed 78 · Task 820/20OOOOOOOOOOOOOOOOOOOONone
PyTorch eagerSeed 79 · Task 919/20OOOOOOOOOOOOOOOOOOOx19
ZETIC optimizedSeed 70 · Task 020/20OOOOOOOOOOOOOOOOOOOONone
ZETIC optimizedSeed 71 · Task 120/20OOOOOOOOOOOOOOOOOOOONone
ZETIC optimizedSeed 72 · Task 220/20OOOOOOOOOOOOOOOOOOOONone
ZETIC optimizedSeed 73 · Task 318/20OOOOOOOOOOOOOxOOOOOx13, 19
ZETIC optimizedSeed 74 · Task 420/20OOOOOOOOOOOOOOOOOOOONone
ZETIC optimizedSeed 75 · Task 520/20OOOOOOOOOOOOOOOOOOOONone
ZETIC optimizedSeed 76 · Task 620/20OOOOOOOOOOOOOOOOOOOONone
ZETIC optimizedSeed 77 · Task 720/20OOOOOOOOOOOOOOOOOOOONone
ZETIC optimizedSeed 78 · Task 820/20OOOOOOOOOOOOOOOOOOOONone
ZETIC optimizedSeed 79 · Task 920/20OOOOOOOOOOOOOOOOOOOONone

O = success · x = failure.

Build with ZETIC

Optimize v0.5-libero.

Share your v0.5-libero checkpoint for AGX Thor, with goals for lower latency, energy per inference and memory use.

We’ll reply by email.