A dexterous simulation platform pairs 50 demonstrated tasks with 16 demonstration-free transfer tasks and execution diagnostics.
NVIDIA Isaac Sim / Isaac Lab with PhysX. Optional Newton evaluator currently covers two tasks.66: Core-50 + Transfer-16
See the geometry. Feel the contact.
Benchmarks and policies for precise, contact-rich robot manipulation.
A 106-task bimanual dexterous suite uses matched in-/out-of-domain variants to evaluate five capability axes across two embodiments.
MuJoCo106 canonical specifications across five capability axes
A printable 36-task suite instantiates eight everyday contact skills with tunable physical/simulation mechanisms and evaluates learned policies on six physical tasks.
PyBullet-linked rigid/constraint mechanism suite with custom tactile rendering36 reference tasks; six physical policy-evaluation tasks
A Core–Gen protocol groups30 dexterous tasks by functional interaction, with limited target demonstrations, variation-controlled rollouts and tactile ablations.
ManiSkill 3 / SAPIEN / PhysX30 tasks in6 families:24 Core+6 Gen. Standard paper comparison evaluates15 tasks in3 families; six scene layouts are not additional task types.
Spatially aligned VTLA pretraining uses human touch as visual supervision and as an optional action-expert input; Spark Arena separates vision-only simulation from real tactile control.
Not reported in the inspected paper; Spark Arena-Sim reproduces the Gen1Prototype kinematics/cameras.12 track-specific task settings: 7 simulation + 5 real; Stack Bowls occurs in both tracks.
Eight simulation tasks and five separate physical tasks compare wrist force integration under a common demonstration, controller and evaluation protocol.
NVIDIA Isaac Lab / Isaac Sim8 simulation tasks; 5 separate real-world tasks. These tracks are not 13 independent task families.
Hierarchical industrial assembly with seven operation types, scoring that accounts for geometric/material/tolerance complexity, and physical plus MuJoCo implementations.
MuJoCo59 assembly primitives spanning 7 operation types and 3 ordered subassemblies; 15 representative task geometries illustrated
Laboratory dexterity benchmark with 14 tool/instrument tasks and controllable dynamic visual phenomena, using hybrid Gaussian-background rendering.
mjlab / MuJoCo Warp with Gaussian-splatting backgrounds14 tasks across 4 robot embodiments; three visual experimental phenomena and six laboratory backgrounds
Paired simulation and physical evaluation across property perception, fragility, sustained contact and precision insertion. The study compares VLA, world-action and tactile-enabled policies.
NVIDIA Isaac Sim / Isaac Lab with modified TacEx / UIPC11 simulation tasks; 8 paired physical tasks
Controlled seven-hand comparison of action representations under shared scenes, task criteria and balanced demonstration budgets.
MuJoCo (DexJoCo-based benchmark)6 tasks × 7 hands = 42 hand–task pairs
Language-conditioned memory benchmark with 90 tasks and explicit cue-hiding design, extending the earlier 32-task RL suite.
ManiSkill / SAPIEN / Gymnasium90 language-conditioned tasks across 10 memory types; 38 Short / 30 Medium / 22 Long
Controlled real-world comparison of tactile encoder and policy-conditioning choices across contact-rich tasks, with over 2,000 rollouts.
Not applicable — real-world controlled study4 physical contact-rich tasks
No matching entries.
Try a task name, another phrase, or fewer filters.
end
The local catalog could not be loaded.
0 selected