LIBERO-Dynamic
LIBERO-Dynamic · Robot manipulation benchmark

Objects keep moving. So should evaluation.

A controlled extension of LIBERO for moving-object manipulation. Study how object motion, inference latency, and action execution shape robot performance.

Code & README Evaluation package Data generation
40LIBERO tasks
1,800episodes / setting
9,000dynamic demos
6pretrained policies
Overview of LIBERO-Dynamic
Object motion and inference latency, evaluated together.
01 / Findings

Main results

Three complementary views of dynamic manipulation.

Cross-model action horizon sweeps

Pretrained policies

All six policies degrade as target speed increases, but their sensitivity and best measured action horizons differ.

Click the figure to enlarge ↗
Aggregate tables are released as CSV files in results/. VLASH uses its own delay-aware checkpoint and is shown as a system reference rather than a same-checkpoint controller ablation.
02 / Benchmark

Explore the 40 tasks

Four LIBERO suites. Three motion patterns.
Examples shown at 0.09 m/s.

T01 · Linear

Put both the alphabet soup and the tomato sauce in the basket

Each demonstration combines a manipulation goal with continuous target motion.

03 / Execution

Execution under inference latency

Same task · successful runs · measured latency

Sync

41 hold frames
Model & execution settings
Model
Pi0.5 PyTorch
Latency
147.2 ms (2.94 ticks @ 20 Hz)
Chunk
K = 7

Async

continuous
Model & execution settings
Model
Pi0.5 PyTorch
Latency
147.2 ms (2.94 ticks @ 20 Hz)
Chunk
K = 6

RTC

continuous
Model & execution settings
Model
Pi0.5 PyTorch RTC
Latency
169.8 ms (3.40 ticks @ 20 Hz)
Chunk
K = 16

VLASH

continuous
Model & execution settings
Model
VLASH continuous
Latency
104.0 ms (2.08 ticks @ 20 Hz)
Chunk
K = 5
Sync, Async, and RTC share the same π0.5 task setup. Sync keeps real inference-hold frames. VLASH uses its own model and measured latency.
04 / Adaptation

After dynamic-data fine-tuning

Successful executions · unseen dynamic phases

Dynamic + static fine-tuning

π0.5

success
L10-03 · Turn on the stove and place the moka potAsync · K=4 · 60.3 ms · 0.09 m/s
Dynamic fine-tuning

History FastWAM-Joint

success
Go-02 · Put the bowl on the stoveAsync · K=8 · 89.2 ms · 0.09 m/s
05 / Real robot

Dynamic tasks in the real world

Four tasks, with selectable example trials.

Task 01

Conveyor-Belt Apple Pick-and-Place

Trial 1 / 2
Task 02

Zigzag Cart Apple Pick-and-Place

Trial 1 / 3
Task 03

Curvilinear Cart Apple Pick-and-Place

Trial 1 / 2
Task 04

Dynamic Mole Hitting

Trial 1 / 1