01 / INTERCEPT
Approach a moving target
The target keeps translating and rotating. The arm must choose where to approach and when to close the gripper.
What to watch Translation and rotation before contact
ASTROBENCH / EMBODIED MANIPULATION
Free-floating
Manipulation Benchmark
AstroBench
How can a robot grasp a drifting, rotating object and keep it secure after contact? AstroBench studies capture and retention, with a broader task design covering mechanisms, assembly and stowage.
Fixed-base arm · Zero-gravity cabin · Free-floating objects
RECORDED EPISODEMoving / Spool
20Object types in the gallery
40Static and moving examples
80Videos across two views
Browse all recordings ↗01 / THE RESEARCH QUESTION
Grasping without support
Without a supporting surface, contact can push an object away or set it spinning. Approach timing, contact formation and retention all matter.
02 / STABILIZE
Manage the first contact
A touch from one finger can deflect the object. Both fingers need to establish contact before it leaves the grasp region.
What to watch First touch through gripper closure
03 / RETAIN
Keep the object secure
Closing the gripper does not guarantee retention. The object can still slip or escape, so its pose relative to the gripper must be tracked.
What to watch Relative motion after closure
Diagrams illustrate the stages. See the recordings for actual episodes.
02 / TASK SYSTEM
Task families
The task design spans 11 families in four categories, from free-floating grasping to interaction with fixed cabin mechanisms.
FREE-FLOATING GRASPING
Grasp a free-floating object
Approach a floating object, close the gripper and maintain a stable grasp. Initial conditions include rest, translation and rotation.
- Constraints
- The target floats freely; the robot base is fixed.
- Challenge
- Target motion, contact timing and stability after contact.
- Completion
- Target-specific bilateral contact, measured closure and continuous retention.
MECHANISM INTERACTION
Operate cabin mechanisms
Open and close doors, move drawers, press buttons, turn knobs or valves, and switch levers.
- Constraints
- Mechanisms are mounted in the cabin; moving parts follow joints or rails.
- Challenge
- Apply force along the permitted motion and reach the target state.
- Completion
- Measured angle or travel, with task-specific release and stage checks.
PRECISION MATING
Insert and assemble
Peg insertion, plug insertion and removal, and module seating.
- Constraints
- A grasped part mates with a hole, socket or guide rail.
- Challenge
- Maintain the grasp while aligning the part and inserting it along the required direction.
- Completion
- Alignment error, insertion depth and seating in the target frame.
These tasks cover geometric insertion and seating, with separate criteria for each task.
OBJECT MANIPULATION
Reorient and stow
Reorient an object, or stow it and close the container.
- Constraints
- Start with a grasped object and change its orientation or place it inside a container.
- Challenge
- Adjust the pose, then complete placement, release and closure in sequence.
- Completion
- The target pose, or an ordered sequence of placement, release, closure and observation.
Stowage requires release and therefore has a different judge from grasp retention.
The current recordings cover free-floating grasping.Task definitions and scope ↗
03 / EVALUATION DESIGN
Evaluation design
The proposed evaluation separates changes in object geometry from changes in motion, then tests both together. Formal splits remain to be frozen.
Geometry × dynamics
Columns: motion conditions · Rows: object identity
| Seen motion | Held-out motion | |
|---|---|---|
| Seen objects | Reference conditionIn-distribution performance | Dynamics generalizationChanged motion |
| Unseen objects | Geometry generalizationChanged objects | Joint generalizationNew objects and motion ↗ |
Related geometry belongs to one split. Multiple episodes of one object retain the same object identity.
Evaluation workflow
- Register tasks and episodesBind assets, geometry groups, initial states, seeds and protocol versions.
- Record executionUse a declared observation and action interface; log states, contacts and timestamps.
- Score and reportCheck continuous retention at the horizon. Report success and completion coverage separately.
PRIMARYStrict task success
DIAGNOSTICSCapture time · Retention · Loss
COMPLETENESSCoverage · Errors · Missing episodes
04 / EPISODE LIBRARY
Episode library
Static and moving grasp examples across 20 object types. Switch cameras within an episode, or expand its record to inspect motion and hold measurements.
Rubik’s cube
Moving targetMotion and hold record
Pre-contact displacement 56.14 mm, speed 6.000 mm/s; pre-contact rotation 29.15°, angular speed 0.0514 rad/s; passed the historical 1-second hold check
Rectangular block
Moving targetMotion and hold record
Pre-contact displacement 53.38 mm, speed 6.000 mm/s; pre-contact rotation 30.63°, angular speed 0.0500 rad/s; passed the historical 1-second hold check
Cylindrical part
Moving targetMotion and hold record
Pre-contact displacement 50.79 mm, speed 6.000 mm/s; pre-contact rotation 29.47°, angular speed 0.0511 rad/s; passed the historical 1-second hold check
Spherical part
Moving targetMotion and hold record
Pre-contact displacement 51.11 mm, speed 6.000 mm/s; pre-contact rotation 29.62°, angular speed 0.0509 rad/s; passed the historical 1-second hold check
Sealed bottle
Moving targetMotion and hold record
Pre-contact displacement 50.93 mm, speed 6.000 mm/s; pre-contact rotation 29.97°, angular speed 0.0506 rad/s; passed the historical 1-second hold check
Round can
Moving targetMotion and hold record
Pre-contact displacement 59.85 mm, speed 6.000 mm/s; pre-contact rotation 28.13°, angular speed 0.0523 rad/s; passed the historical 1-second hold check
Sample tube
Moving targetMotion and hold record
Pre-contact displacement 53.39 mm, speed 6.000 mm/s; pre-contact rotation 29.94°, angular speed 0.0506 rad/s; passed the historical 1-second hold check
Lidded box
Moving targetMotion and hold record
Pre-contact displacement 51.25 mm, speed 6.000 mm/s; pre-contact rotation 29.67°, angular speed 0.0509 rad/s; passed the historical 1-second hold check
Screwdriver
Moving targetMotion and hold record
Pre-contact displacement 51.67 mm, speed 6.000 mm/s; pre-contact rotation 29.11°, angular speed 0.0514 rad/s; passed the historical 1-second hold check
Wrench
Moving targetMotion and hold record
Pre-contact displacement 52.40 mm, speed 6.000 mm/s; pre-contact rotation 30.61°, angular speed 0.0507 rad/s; passed the historical 1-second hold check
Pliers
Moving targetMotion and hold record
Pre-contact displacement 57.13 mm, speed 6.000 mm/s; pre-contact rotation 32.28°, angular speed 0.0489 rad/s; passed the historical 1-second hold check
Small hammer
Moving targetMotion and hold record
Pre-contact displacement 69.92 mm, speed 6.000 mm/s; pre-contact rotation 30.02°, angular speed 0.0504 rad/s; passed the historical 1-second hold check
Flashlight
Moving targetMotion and hold record
Pre-contact displacement 58.00 mm, speed 6.000 mm/s; pre-contact rotation 30.14°, angular speed 0.0504 rad/s; passed the historical 1-second hold check
Handheld terminal
Moving targetMotion and hold record
Pre-contact displacement 85.36 mm, speed 6.000 mm/s; pre-contact rotation 42.85°, angular speed 0.0389 rad/s; passed the historical 1-second hold check
Small camera
Moving targetMotion and hold record
Pre-contact displacement 48.56 mm, speed 6.000 mm/s; pre-contact rotation 28.83°, angular speed 0.0517 rad/s; passed the historical 1-second hold check
Spool
Moving targetMotion and hold record
Pre-contact displacement 46.83 mm, speed 6.000 mm/s; pre-contact rotation 27.57°, angular speed 0.0528 rad/s; passed the historical 1-second hold check
L-bracket
Moving targetMotion and hold record
Pre-contact displacement 46.10 mm, speed 6.000 mm/s; pre-contact rotation 28.13°, angular speed 0.0523 rad/s; passed the historical 1-second hold check
T-fitting
Moving targetMotion and hold record
Pre-contact displacement 57.58 mm, speed 6.000 mm/s; pre-contact rotation 31.94°, angular speed 0.0486 rad/s; passed the historical 1-second hold check
Battery pack
Moving targetMotion and hold record
Pre-contact displacement 55.90 mm, speed 6.000 mm/s; pre-contact rotation 31.82°, angular speed 0.0488 rad/s; passed the historical 1-second hold check
Ring part
Moving targetMotion and hold record
Pre-contact displacement 60.31 mm, speed 6.000 mm/s; pre-contact rotation 33.53°, angular speed 0.0472 rad/s; passed the historical 1-second hold check
Rubik’s cube
Initially staticMotion and hold record
Passed the original grasp check · Two views recorded at 30 Hz
Rectangular block
Initially staticMotion and hold record
Passed the original grasp check · Two views recorded at 30 Hz
Cylindrical part
Initially staticMotion and hold record
Passed the original grasp check · Two views recorded at 30 Hz
Spherical part
Initially staticMotion and hold record
Passed the original grasp check · Two views recorded at 30 Hz
Sample tube
Initially staticMotion and hold record
Passed the original grasp check · Two views recorded at 30 Hz
Sealed bottle
Initially staticMotion and hold record
Passed the original grasp check · Two views recorded at 30 Hz
Round can
Initially staticMotion and hold record
Passed the original grasp check · Two views recorded at 30 Hz
Lidded box
Initially staticMotion and hold record
Passed the original grasp check · Two views recorded at 30 Hz
Screwdriver
Initially staticMotion and hold record
Passed the original grasp check · Two views recorded at 30 Hz
Wrench
Initially staticMotion and hold record
Passed the original grasp check · Two views recorded at 30 Hz
Pliers
Initially staticMotion and hold record
Passed the original grasp check · Two views recorded at 30 Hz
Small hammer
Initially staticMotion and hold record
Passed the original grasp check · Two views recorded at 30 Hz
Flashlight
Initially staticMotion and hold record
Passed the original grasp check · Two views recorded at 30 Hz
Battery pack
Initially staticMotion and hold record
Passed the original grasp check · Two views recorded at 30 Hz
Handheld terminal
Initially staticMotion and hold record
Passed the original grasp check · Two views recorded at 30 Hz
Small camera
Initially staticMotion and hold record
Passed the original grasp check · Two views recorded at 30 Hz
Ring part
Initially staticMotion and hold record
Passed the original grasp check · Two views recorded at 30 Hz
Spool
Initially staticMotion and hold record
Passed the original grasp check · Two views recorded at 30 Hz
L-bracket
Initially staticMotion and hold record
Passed the original grasp check · Two views recorded at 30 Hz
T-fitting
Initially staticMotion and hold record
Passed the original grasp check · Two views recorded at 30 Hz
No matching objects. Try another search.
About the recordings
Moving examples begin with translation and rotation; static examples begin at rest without support. Both cameras show the same episode and preserve playback position when switched. Videos end with about one second of frozen footage; hold labels come from the corresponding logs.
Recording-era hold labels and the current terminal-retention protocol use different time criteria. The specification explains both.
05 / DOCUMENTATION
Technical specification
Task definitions, physical assumptions, interfaces, scoring, reproducibility and the matched cube pilot.
Code, paper and evaluation status
The evaluation protocol is available on GitHub. A 40-episode matched cube pilot is described in the specification. Environment packaging and formal splits are still in progress. The AstroBench paper is available as arXiv:2610.31064 ↗.
06 / PAPER
Citing AstroBench
The benchmark is described in an arXiv preprint. Please cite it when using AstroBench tasks, objects or evaluation protocols.
AstroBench: Benchmarking Free-Floating Capture and Space Robotic Manipulation
@article{wang2026astrobench,
title = {AstroBench: Benchmarking Free-Floating Capture and Space Robotic Manipulation},
author = {Wang, Zizhe and Deng, Jia and Lu, Xin and Jia, Zehao and Huang, Yi and Zhu, Yao},
journal = {arXiv preprint arXiv:2610.31064},
year = {2026}
}