ContactBench 1.0: A Benchmark for Everyday Contact-Rich Robotic Manipulation

Jingyi Xiang; Yixuan Wang; Jeannette Bohg; Wenzhen Yuan

cs.ROOct 17, 2026

ContactBench 1.0: A Benchmark for Everyday Contact-Rich Robotic Manipulation

Authors: Jingyi Xiang, Yixuan Wang, Jeannette Bohg, Wenzhen Yuan

Organizations: University of Illinois Urbana-Champaign, Urbana, IL, USA · Stanford University, Palo Alto, CA, USA

Abstract

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.

Figures & tables

Fig. 2: The eight contact-rich manipulation skills we identified for everyday contact-rich manipulation tasks. We built ContactBench 1.0 to instantiate these skills.

Fig. 3: ContactBench 1.0 includes 36 manipulation tasks to concretely evaluate a robot system’s proficiency on the eight skills. Many tasks require more than one skill to complete. We record humans solving the tasks to illustrate successful task executions due to the difficulty of many tasks for current robot systems.

Task	Variations
Turn doorknob to open door Turn door handle to open door	Required rotation amount
	Rotation force
	Object geometry
Plug in charger	Insertion force
Plug in USB	Object geometry
Put bottle cap onto bottle	Required rotation amount Object geometry

TABLE I: Task variations supported by our benchmark. Blue rows denote behavioral variations that affect task behavior but are ambiguous to vision; green rows denote visual + behavioral variations that affect both task behavior and appearance. Terminology follows Star-Gen [ 49 ] .

Fig. 4: We design spring structures inside the assets to approximate their resistance force during manipulation. Here is an example of the torsional spring inside a door handle and a compression spring inside an outlet.

Fig. 5: The resistance forces exhibited by our assets are comparable to those of real objects and are generally consistent across different 3D printers. We show experimental results over a door handle and a charger-outlet pair with their variants. (a) shows the experimental setup, (b) shows the real products used for comparisons, and (c) shows the measured max pressing and insertion forces across the real objects and the asset variants printed by two printers.

Fig. 6: Overview of the ContactBench 1.0 simulation suite. The suite provides assets that match the 3D-printable versions in terms of shape and mechanisms (a) , a simulated GelSight Mini interface (b) , and a teleoperation interface for demonstrating manipulation policies through keyboard, gamepad, or SpaceMouse (c) . We also provide a set of scripted policies for the contact-rich manipulation tasks.

Fig. 7: Our simulated asset models feature tunable force response. In this example of the door handle asset, we make a robot rotate it at a linear speed of 10 mm/s. Here we show the measured force/torque value at the robot wrist. (a) Force response when changing the simulated door handle’s parameter of revolute spring with stiffness $K$ . (b) Force response when changing the simulated door handle’s parameter of pre-load torque $\tau_{0}$ . (c) Force response of simulated assets with varied settings in comparison with the response from real assets. The real ones are printed on a Bambu Lab P1S with 5, 7.5, and 10 mm springs (solid, mean of 10 trials). The simulation reproduces the shape and ordering of the measured profiles.

Fig. 8: Example of the scripted policy in simulation with tactile feedback: key insertion and turning. The correctly oriented key is inserted and turned to retract the bolt (1–3). The keyway interlock rejects a flipped-key insertion attempt (4). Insets show simulated tactile images from the right fingertip; yellow arrows indicate marker displacement.

Panel of Figure 9 — Fig. 9: Example of the scripted policy in simulation: push-and-turn dial operation. The plots show fingertip forces, dial rotation, and axial depression during the process, with numbered guides matching the keyframes. The red, orange, and green bands denote the phase of blocked turning, pressing, and permitted rotation, respectively.

Task Execution Sequence Using UMI	Task Specs
	Task: Thread nut onto screw Skills: Insertion and alignment, articulated object manipulation Description: Screw the nut held in hand onto the screw head placed in the workspace such that their threads fully engage. Success criteria: The nut and screw threads engage. The robot movement stops after terminating the rotation motion, moving away from, and breaking contact with task objects.
	Task: Tighten screw with Allen key Skills: Extrinsic contact control, insertion and alignment Description: A screw is placed into the screw slot of an assembly block in the workspace, but the threads are unengaged. Use the Allen key grasped in hand to tighten the screw. Success criteria: The screw is tightened and remains stationary when a small force is applied in an attempt to rotate the screw. The robot movement stops after terminating the tightening motion, moving away from, and breaking contact with the screw.
	Task: Plug in charger Skills: Insertion and alignment Description: An outlet is placed in the workspace. Fully insert the charger held in hand into the outlet. Success criteria: The charger is fully inserted into the outlet, such that it does not noticeably move when pushed further toward the outlet. The robot movement stops after terminating the insertion motion, moving away from, and breaking contact with task objects.
	Task: Turn doorknob to open door Skills: Articulated object manipulation Description: Turn the doorknob fully to unlock the door, then open the door. Success criteria: The door is opened by at least 5 degrees, such that it is visually obvious that the door has been unlocked.
	Task: Retrieve single utensil from bin Skills: Grasping from clutter, grasping thin or small objects Description: A utensil storage bin is placed in the workspace. The target utensil is placed in the bin alongside multiple other utensils. Pick up a single copy of the target utensil from the cluttered storage bin and place it on the table without removing other objects from the bin. Success criteria: The robot movement stops after separating a single copy of the target utensil and placing it on the table. Other utensils should remain in the storage bin. The robot should not remain in contact with the task objects.
	Task: Pick up card from platform Skills: Grasping thin or small objects Description: Pick up the thin and/or small object placed on the flat platform. Success criteria: The target object is stably grasped by the robot and lifted off the platform.

TABLE II: Benchmark tasks used for policy learning experiments

Human

RDP

\pi_{0.5}

FTP-1

Task

Time

\uparrow

Prog

\uparrow

>

40 N

\downarrow

Time

\downarrow

\uparrow

Prog

\uparrow

>

40 N

\downarrow

Time

\downarrow

\uparrow

Prog

\uparrow

>

40 N

\downarrow

Time

\downarrow

\uparrow

Prog

\uparrow

>

40 N

\downarrow

Time

\downarrow

Thread nut onto screw

17.5 s

0 8/30

0.39

33%

1.72

\times

0 0/30

0.24

27%

–

14/30

0.66

27%

1.22

\times

14/30

0.63

83%

1.32

\times

Tighten screw with Allen key

19.7 s

0 0/30

0.58

10%

–

0 0/30

0.36

–

17/30

0.82

1.50

\times

0 3/30

0.45

0.84

\times

†

Plug in charger

7.3 s

0 2/30

0.32

80%

1.06

\times

†

0 4/30

0.34

83%

1.15

\times

†

0 6/30

0.44

40%

0.89

\times

0 4/30

0.35

40%

0.95

\times

†

Turn doorknob to open door

7.6 s

30/30

1.00

10%

0.92

\times

29/30

0.98

63%

0.91

\times

30/30

1.00

0.87

\times

29/30

0.98

43%

0.89

\times

Retrieve single utensil from bin

11.8 s

0 0/30

0.50

97%

–

not evaluated

16/30

0.70

47%

2.05

\times

not evaluated

TABLE III: Policy performance on the six selected tasks (30 episodes per method per task)

Fig. 10: We systematically sample task object initial positions within a specified region and render them in a 3rd-person camera view to guide the human operator in task object placement. When necessary, the task objects are attached to the workspace with Velcro.

Fig. 11: Task progress achieved by each policy, averaged over 30 evaluation trials.

Fig. 12: Examples of the four typical challenges vision-only policies encountered on the selected tasks.

Fig. 13: Top: Both RDP and FTP-1 are prone to terminating the tightening motion without tightening the screw after the Allen key slips in hand. Bottom: Although in-hand slip can happen naturally as the task progresses, the correct tactile signal associated with a tightened screw has a large in-plane torque.

Fig. 14: Left: Success rates across task variants for turn doorknob to open door . Right: Robot wrist rotation angle distribution for each policy. Only RDP shows behavior adaptation effort when the task demands larger rotation angles to unlock the door.

Variant	DP	RDP	$\pi_{0.5}$	FTP-1
60 mm knob (train)	$58.9\pm 0.12$	$59.0\pm 0.13$	$58.8\pm 0.13$	$59.2\pm 0.14$
55 mm knob	$59.0\pm 0.11$	$59.0\pm 0.11$	$58.7\pm 0.11$	$58.9\pm 0.14$

TABLE IV: Gripper width reached (mm) while trying to turn the doorknob, reported as mean

\pm

std across episodes (30 episodes for the 60 mm knob, 10 for the 55 mm knob).

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Fig. A1: Examples of CAD-derived simulation assets. The assets approximate the corresponding physical mechanisms using rigid parts, joints, and simplified representations of compliance and mechanical constraints. Exploded views of (a) a push-and-turn dial, (b) a door handle, (c) a key-operated lock, (d) a doorknob, and R and P denote revolute and prismatic degrees of freedom, respectively.

Fig. A2: Simulated GelSight Mini responses to a key under normal loading, shearing, and rotation. Top: rendered tactile images, with red arrows indicating marker displacement. Bottom: gel indentation height maps on a common color scale.

Fig. B1: Our setup for demo collection and policy rollout.

	DP	RDP	$\pi_{0.5}$	FTP-1
Initialization	From scratch	From scratch	pi05_base , full fine-tuning	ftp1_v0426_50kstep , full fine-tuning
Data rate	12 Hz	12 Hz (slow policy), 24 Hz (fast policy)	24 Hz	24 Hz
Visual input	Wrist RGB, $240\times 320$	Wrist RGB, $240\times 320$	Wrist RGB, $240\times 320$ , resized then zero-padded to $224\times 224$	Wrist RGB, $240\times 320$ , resized then zero-padded to $224\times 224$
Tactile input	–	$2\times$ GelSight Mini marker movements, each projected to 15-D with the PCA basis released with RDP	–	$2\times$ GelSight Mini images, each resized to $224\times 224$ ; tactile encoder initialized from FTP-1’s pretrained GelSight Mini checkpoint
Proprioceptive input	EEF pose (relative to current), absolute gripper width (normalized)	EEF pose (relative to current), absolute gripper width (normalized)	Absolute gripper width (normalized); EEF pose set to identity	Absolute gripper width; EEF pose set to identity
Observation history	2 steps	2 steps	1 step (current only)	1 step (current only)

Appendix

TABLE B1: Training details of the evaluated policies.

Task	Demos	Mean (s)	Min–max (s)	Language instruction
Thread nut onto screw	100	19.6	14.0–24.0	“screw the nut onto the screw”
Tighten screw with Allen key	100	21.0	11.9–32.7	“tighten the screw with allen key”
Plug in charger	100	8.4	7.0–10.0	“plug the charger into the outlet”
Turn doorknob to open door	99 ∗	8.8	7.5–11.0	“turning the door knob to open the door”
Retrieve single utensil from bin	100	12.9	8.1–27.5	“pick up a single utensil from the storage bin”
Pick up card from platform	100	8.6	7.0–10.5	“pick up the thin object on the table”

Appendix

TABLE B2: Demonstrations used for training and language instructions given to

\pi_{0.5}

and FTP-1.

Contact-Rich Robotic Manipulation

Contact-rich robotic manipulation studies how robots plan and control physical interactions such as grasping, insertion, pushing, and dexterous object handling.

See all 413 papers

IMBench: A Benchmark for Intuitive Robotic Manipulation
Jul 17, 2026Anurag Maurya, Sukhvansh Jain, Prajwal Avhad +9Contact-Rich Robotic Manipulation Robotic Manipulation
Manav Robotics, Bengaluru, India
PredTac: Learning Contact-Rich Manipulation with Predicted Touch
Sep 14, 2026Weijia Fan, Daqiang GuoVisual-to-Tactile Prediction Tactile Sensing
The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China.
Real-IKEA: Physical Fidelity is the Prerequisite for Robust Manipulation
Jun 7, 2026Kunqi Xu, Zhenhao Huang, Siyuan Luo +2Contact-Rich Robotic Manipulation Robotics Simulation
Human-centered Robotic Lab, National University of Singapore · School of EECS, Peking University

Jul 17, 2026J/K move · Enter open · S save

IMBench: A Benchmark for Intuitive Robotic Manipulation

Anurag Maurya, Sukhvansh Jain, Prajwal Avhad, Gautham Balachandran+8

Manav Robotics, Bengaluru, India

Contact-Rich Robotic Manipulation Robotic Manipulation Robot Manipulation Benchmarks VLMs for Robotics

Open paper AskPDF arXiv

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.

Abstract

Contact-Rich Robotic Manipulation

Explore similar work

IMBench: A Benchmark for Intuitive Robotic Manipulation

PredTac: Learning Contact-Rich Manipulation with Predicted Touch

Real-IKEA: Physical Fidelity is the Prerequisite for Robust Manipulation