Capturing a loose object inside a spacecraft requires a robot to approach a target that can translate and rotate, and whose motion changes when contact is made. These conditions motivate evaluation beyond manipulation of supported objects. We introduce AstroBench, an open benchmark for free-floating capture and space robotic manipulation with a fixed-base arm. The task suite combines four initial-motion tracks with sample-storage operations based on MELFI hardware and cap and captive-screw servicing based on the Robotic Refueling Mission. The two workcells provide nine task definitions and 29 task-target combinations. A metadata-filtered Objaverse inventory supplies 8,414 object instances for asset expansion, alongside five cabin visual contexts and versioned task instructions. We report two acquisition-focused experiments using existing checkpoints. A frozen π0.5 DROID joint-position policy captures 2 of 200 static targets and 4 of 200 moving targets across twenty object identities. A cube-specific ACT policy captures 8 of 10 static targets and 6 of 10 moving targets in a separate experiment. Each experiment retains its recorded observation, action and capture configuration. The results document difficult direct transfer for the tested generalist checkpoint and successful task-specific capture behavior, while identifying concrete settings for studying motion-aware perception and contact-sensitive control. Code, task definitions and editable workcell assets are available at https://github.com/AstroUpLab/AstroBench.
Figures & tables
Fig. 1: AstroBench combines free-floating capture, MELFI sample-storage operations and RRM cap and fastener servicing. Panel (a) is a recorded π0.5 rollout; panels (b) and (c) show the current workcell geometry. The capture target is unsupported; the storage and service hardware is secured to its worksite. The RRM panel is wall-mounted, with a floor-mounted arm centered in front of it.
Benchmark
Emphasis relevant to this work
LIBERO
Transfer across structured manipulation suites
CALVIN
Language-conditioned, long-horizon composition
RoboCasa
Diverse everyday tasks and visual environments
Space Robotics Bench
Mission-relevant robot learning in space environments
AstroBench
Free-floating capture and reference-based storage/service interfaces
TABLE I: Research emphasis of representative related benchmarks. This comparison describes the questions addressed rather than ranking platform performance.
Track
Linear speed (m/s)
Angular speed (rad/s)
Static
0
0
Drift
0.02
0
Tumble
0.02
0.5
Fast tumble
0.05
1.0
TABLE II: Initial-motion tracks in the released capture catalog. Magnitudes are shown; registered seeds determine the signed directions.
Workcell
Task
Targets
Initial condition
Goal evidence
MELFI
Card transfer
3
Card outside; gripper open
Card acquired and aligned with the selected opening
Cap removed with its adapter; assembly docked; driver withdrawn
TABLE III: Workcell tasks and their operation goals. The target count expands each task definition into concrete language-conditioned cases.
Setting
Frozen π0.5 DROID
Cube-specific ACT
Training use
Existing checkpoint; no adaptation in this evaluation
400 local cube demonstrations; trained from random initialization; frozen at step 140,000
Objects
Twenty authored object identities
One Rubik’s cube
Initial states
One nominal pose per object, repeated ten times per condition
Ten fixed development positions, each evaluated once per condition
Conditions
Static / translation plus rotation
Static / translation plus rotation
Moving-target initialization
0.006 m/s along world x ; 0.08 rad/s about world z
0.006 m/s along world x ; 0.08 rad/s about world z
Physics / action rate
120 Hz / 15 Hz
120 Hz / 15 Hz
TABLE IV: Recorded capture configurations. Each column defines its own experiment; object coverage, initial states, detectors and action scheduling differ.
Condition
Attempts
Captures
Rate
Rejections
Static
200
2
1.0%
0
Dynamic
200
4
2.0%
2
Total
400
6
1.5%
2
TABLE V: π0.5 acquisition results. Rejections are included in the attempt denominator and shown separately.
Condition
Attempts
Captures
Rate
Static
10
8
80%
Moving
10
6
60%
Total
20
14
70%
TABLE VI: Cube-specific ACT acquisition results in the 20 s experiment.
Fig. 2: Acquisition measurements from the two recorded experiments. The π0.5 panel resolves the twenty-object evaluation by object and motion condition; each cell has ten attempts. ACT is shown in a separate panel for its cube-specific experiment. Table IV specifies the settings of each panel.
Fig. 3: Recorded ACT capture of an initially stationary cube, shown at 0, 4, 8 and 10 s in one of the twenty evaluation attempts. The measured capture predicate first qualifies at 8.07 s. Frames illustrate approach and acquisition; the event time is obtained from the synchronized state trace rather than inferred from these sparsely sampled images.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Captures / attempts
Rej.
Object
Static
Dynamic
S
D
Cube
0/10
0/10
0
0
Rectangular block
0/10
0/10
0
0
Cylinder
0/10
0/10
0
1
Sphere
0/10
0/10
0
0
Sample tube
1/10
0/10
0
0
Appendix
TABLE VII: π0.5 per-object capture counts. Each motion condition has ten attempts per object. Rej. counts predicted joint-limit rejections, included in the attempt denominators.
Label
Task ID
Exact English prompt template
M1
card_transfer_to_mouth
Move the card to the opening of the {target_name} and hold it aligned.
M2
card_insert
Insert the card fully into the {target_name}, release it, and leave it retained inside.
M3
card_extract
Remove the card from the {target_name} and hold it clear of the opening.
M4
lid_open
Open the lid of the {target_name} and leave it open.
M5
lid_close
Close and latch the lid of the {target_name}, then withdraw the gripper.
R1
fdv1_remove_stow
Unscrew and remove the {target_name}, retain it in its adapter, dock the adapter in its stowage holder, and withdraw the driver.
Appendix
TABLE VIII: Canonical workcell instruction templates. Resolve each target placeholder using the exact names in Table IX . The labels M1–R4 are local table abbreviations; task IDs are the machine-readable registry keys.
Templates
Target ID
Exact target_name
M1–M5
Front_quarter_box
front quarter box
M1–M5
Middle_quarter_box
middle quarter box
M1–M5
Rear_half_box
rear half box
R1
FDV1
FDV1 cap
R2
SMA_1_1
SMA cap at row 1, column 1
R2
SMA_1_2
SMA cap at row 1, column 2
Appendix
TABLE IX: Exact target IDs and instruction names. Each row is paired with every listed template, giving 29 distinct task–target combinations.
Fig. 4: MELFI storage hardware. (a) The mixed tray contains two quarter-size modules and one half-size module, with separate lid and tray locks. (b) An open quarter box exposes the hinge, curved perforated shell and released card-restraint band. (c) The sample card carries four vials below an unobstructed grasp region. (d) The half-box extraction reset places the card inside with the band routed around it. Panels (b) and (d) show task reset states. The historical mixed-box arrangement follows ESA and NASA documentation; the card’s 100×110 mm planar dimensions follow ESA’s description [ 11 , 12 ] .
Fig. 5: RRM panel and tool interfaces in the accepted wall-mounted layout. (a) The full board locates the FDV1 interface, SMA array, captive fasteners and the two receivers below the panel. (b) The FDV1 cap and Tertiary Cap Adapter are shown in the centered ready configuration. (c) The 3×3 SMA array supplies addressed cap targets. (d) The two captive-screw positions expose their drive interfaces. (e) The larger TCA receiver and smaller SMA-adapter receiver contain separate latch components. The views use the same final scene and part poses; the arm and waiting tool are hidden where they obstruct the hardware. NASA sources establish the reference panel and tool roles [ 14 , 13 , 15 ] .
Contact-rich manipulation underlies many everyday human activities, yet evaluating robots on such tasks remains difficult: the tasks involve diverse contact behaviors and are hard to reproduce across laboratories. Drawing on observations of everyday human manipulation, we identify eight essential skills that provide a practical framework for studying contact-rich manipulation. We introduce ContactBench 1.0, an extensible benchmark that instantiates these skills through 36 reference tasks and more than 40 object assets. The assets capture task-relevant geometry and contact behavior, can be fabricated reproducibly on consumer-grade 3D printers, and are each paired with a matching simulation model. Controlled task variations in both the real world and simulation enable evaluation of how well robot systems generalize. We evaluate recent representative vision-only and visual-tactile policy learning methods on six benchmark tasks. The results reveal that while vision-only policies struggle to perceive or react to contact events, visual-tactile policies do not meaningfully improve their performance, highlighting that contact-rich manipulation remains a significant challenge for current policy learning methods. By providing common task definitions alongside physical and simulated assets, ContactBench 1.0 offers a shared testbed for developing, evaluating, and comparing contact-rich manipulation systems. The benchmark will be open-sourced at https://jingyi-xiang.github.io/contactbench.
Jingyi Xiang, Yixuan Wang, Jeannette Bohg +1
University of Illinois Urbana-Champaign, Urbana, IL, USA · Stanford University, Palo Alto, CA, USA
Frontier multimodal foundation models (e.g., GPT-6 Astra) have recently shown strong potential for direct robotic control, yet their performance on fine manipulation remains limited. We argue that an important source of failure is not necessarily insufficient policy capability, but insufficient spatial observability, where task-critical spatial relationships may be poorly revealed by the existing physical camera setup. We introduce SpatialHarness, a test-time embodied harness that provides test-time spatial scaffolding for fine robotic manipulation without policy fine-tuning or changes to the physical sensing setup. SpatialHarness maintains an online simulated scene synchronized with real-world execution, identifies task-critical spatial relationships, and renders complementary virtual views that expose them to a frozen multimodal policy. To keep the simulated scene aligned during interaction, we develop interaction-aware scene synchronization that distinguishes static, held, and transition modes. We evaluate SpatialHarness on four real-robot manipulation tasks spanning precise geometric alignment, object-relative placement, and articulated-object interaction. Using the same frozen GPT-6 Astra policy, SpatialHarness substantially improves task success, including from 26.7% to 66.7% on plug insertion and from 0% to 100% on Tower of Hanoi. These results indicate that improving spatial observability at test time can unlock fine-manipulation capabilities already present in strong multimodal foundation models. Project website: https://emilia113.github.io/SpatialHarness/.
Jiayu Wang, Yue Yu, Bin Zhu +2
College of Computer Science and Artificial Intelligence, Fudan University · Singapore Management University · Institute of Trustworthy Embodied AI, Fudan University
Humans combine reasoning and motor control to solve complex manipulation tasks under diverse constraints. They build an understanding of the physical world that helps them convert reasoning into actions and quickly adapt to new scenes, tasks, and rules. We refer to this capability as intuitive manipulation. Existing benchmarks fail to capture this integration: they evaluate physical reasoning in isolation from execution, or measure policy performance without requiring explicit reasoning. We introduce IMBENCH, a benchmark designed to evaluate intuitive manipulation as an integrated capability spanning perception, physical reasoning, action generation, and iterative execution. Our tasks require models to infer task-relevant physical structure and generate feasible action sequences under explicit constraints, including contact-rich manipulation, tool use, and multi-stage dependencies. We introduce a benchmark of 35 tasks, 14K filtered trajectories, and scalable tools for generating diverse scenarios. Experiments reveal a consistent gap: vision language models show partial physical reasoning ability but fail to produce executable plans, while state-of-the-art vision-language-action models struggle to satisfy task constraints and generalize across scenarios. These results identify intuitive manipulation as a missing axis in current foundation models and generalist robot policies, and position IMBENCH as a step toward evaluating and enabling more integrated, adaptive physical intelligence.