CoRL 2026 Conference on Robot Learning

HARP-Bench

Robot manipulation around active humans.

Qinhong Zhou1,* Yun Zong2,* Chunru Lin1 Diyi Hu4 Shivraj Singh Bhatti1 Hongxin Zhang1 Lin Qiu3 Hanqing Zeng3 Chuang Gan1

1University of Massachusetts Amherst 2University of Minnesota Twin Cities 3Meta 4Independent Researcher

*Equal contribution

Four conditions of human involvement

Four ways humans shape the task

Co-presence

The robot works beside a person.

Example shown: the person wipes the table while the robot puts a cube in the cabinet.

Contribution

Human and robot share the work.

Example shown: the person places one piece of bread, and the robot adds the other.

Interruption

The robot waits, then resumes.

Example shown: the person inspects a shoe; the robot waits, then puts it in the box.

Cueing

Human action sets the goal.

Example shown: after the person drinks, the robot pours water into their cup.

Results

37-task evaluation

Success rate (%) on the evaluation seeds of all 37 tasks (up to 100 seeds per task). The full-state expert reads the human's intent from simulator state; the no-intent expert runs the same script without it.

Category Expertfull state Expertno intent ACTDPπ0SmolVLA
Cueing 7 tasks99.038.311.316.916.011.5
Contribution 10 tasks98.853.83.74.216.51.2
Interruption 10 tasks99.80.38.85.79.42.9
Co-presence 10 tasks98.897.418.113.032.69.1
All tasks (37)99.148.210.49.418.85.7

best learned policysecond bestExperts are shown for reference and are not ranked. Averages are unweighted means over tasks.

Per-task results (37 tasks)
Task Expertfull state Expertno intent ACTDPπ0SmolVLA
Cueing (7)
Deliver to Human100.030.20.00.00.00.0
Oil Bottle Recovery95.11.20.08.67.48.6
Open Microwave100.0100.020.037.039.045.0
Pour Water99.090.859.270.463.326.5
Stamp Documents100.014.40.00.00.00.0
Take from Human98.931.60.02.12.10.0
Take from Human (Safety)100.00.00.00.00.00.0
Average99.038.311.316.916.011.5
Contribution (10)
Categorize Objects100.047.50.00.00.00.0
Dump Bin98.441.34.81.612.90.0
Place Bread in Basket100.053.20.01.11.10.0
Place Burger and Fries100.049.50.012.137.41.0
Place Food in Skillet95.319.811.62.329.82.3
Place Pair of Shoes98.743.07.66.315.22.5
Put Object in Cabinet95.590.95.711.444.34.5
Rank Blocks by Color100.085.71.00.010.22.0
Rank Blocks by Size100.073.15.40.011.80.0
Stack Three Bowls100.034.01.07.02.00.0
Average98.853.83.74.216.51.2
Interruption (10)
Categorize Objects100.00.01.02.00.00.0
Dump Bin97.80.01.14.516.30.0
Place Bread in Basket100.01.015.214.19.10.0
Place Burger and Fries100.00.00.01.03.01.0
Place Food in Skillet100.00.050.628.66.07.1
Place Pair of Shoes100.00.015.01.016.012.0
Put Object in Cabinet100.00.03.20.010.52.1
Rank Blocks by Color100.02.00.00.00.00.0
Rank Blocks by Size100.00.00.00.01.10.0
Stack Three Bowls100.00.02.06.032.07.0
Average99.80.38.85.79.42.9
Co-presence (10)
Categorize Objects98.9100.022.625.815.15.4
Dump Bin92.288.95.210.432.30.0
Place Bread in Basket100.0100.04.010.737.30.0
Place Burger and Fries100.0100.00.014.045.25.4
Place Food in Skillet100.0100.048.733.855.67.3
Place Pair of Shoes100.0100.010.05.026.70.0
Put Object in Cabinet96.793.38.910.17.84.4
Rank Blocks by Color100.097.30.00.08.20.0
Rank Blocks by Size100.094.54.70.04.70.0
Stack Three Bowls100.0100.077.320.593.268.2
Average98.897.418.113.032.69.1
All tasks (37)99.148.210.49.418.85.7

Safety metrics

Task success requires safe interaction

We monitor robot–human collisions and minimum separation throughout each episode. Contact with a person marks the episode as a failure; a configurable safety margin flags when the robot gets too close.

The line shows minimum gripper–human separation: green at or above 10 cm, yellow below 10 cm, and red on contact. The red ring marks the collision, which ends the episode as a failure.

Safety evaluation results

Reported success requires task completion without robot–human collision.

PolicySuccess ↑%Collision ↓% episodesWithin 10 cm ↓% episodesTime < 10 cm ↓% of episodeMin. distance ↑cm, mean
ACT10.410.249.38.013.3
DP9.417.352.89.412.0
π018.820.759.98.79.5
SmolVLA5.721.660.29.310.4
Expert full state99.10.038.55.715.8
Expert no intent48.229.964.89.412.6

best learned policysecond best↑ higher is better · ↓ lower is better. Experts are reference baselines and are not ranked.

Collision rate (% of episodes) by condition. Lower is better.

ConditionACTDPπ0SmolVLAExpertfull stateExpertno intent
Cueing13.032.49.629.00.08.5
Contribution7.28.78.56.60.010.1
Interruption14.521.655.341.20.092.8
Co-presence6.811.26.011.80.01.6
All tasks10.217.320.721.60.029.9

Within 10 cm: the robot came within 10 cm of the person at any point. Minimum distance: closest robot–human distance in an episode; contact can make it negative.

Interactive demo

Try controlling the robot
Citation
@inproceedings{harpbench2026,
  title     = {{HARP-Bench}: Benchmarking Robot Manipulation Around Active Humans},
  author    = {Qinhong Zhou and Yun Zong and Chunru Lin and Diyi Hu and Shivraj Singh Bhatti and Hongxin Zhang and Lin Qiu and Hanqing Zeng and Chuang Gan},
  booktitle = {Conference on Robot Learning (CoRL)},
  year      = {2026},
  url       = {https://github.com/UMass-Embodied-AGI/harp}
}