Skip to content

About RoboWits

30 seed tasks on a fixed bimanual robot (two 7-DoF arms, two-finger grippers) in Genesis, plus 163 published mutations that vary the scene while keeping the parent task's success predicate. The premise is that a task should need an insight, not a longer motion: align three cubes by pushing them against a ruler, sweep a screw into a dustpan, pour a ball through a funnel, reach a roll with a rod.

All 30 tasks Capability coverage

Why we are looking at it

It is the opposite end of the axis from BEHAVIOR-1K. There is no navigation, no house, and an episode is seconds rather than minutes — so a task fails for one legible reason, which makes it a good instrument for a coding agent. Several tasks are unsolvable by pick-and-place no matter how precise the arm is, which is exactly the property we want to measure.

The embodiment
Robot Marvin, bimanual — 2 × 7-DoF arms, 2 × 2-finger Pika grippers (18 DoF)
Base Fixed at a table; no navigation
Control EE_ABS (default), EE_DELTA, JOINT_ABS, JOINT_DELTA at 30 Hz
Cameras ego + both wrists, 848 × 480
Privileged state Available in simulation; a task can be run state-based or from pixels

How a task defines success

Every environment class carries its own _check_success, and states the criteria in its docstring. Those criteria are mirrored verbatim into each task page's upstream: block — that list is the benchmark's definition, not our paraphrase. 14 stack cubes, for example:

Base cubes are on the table and touching with ≥ 50% face overlap · Apex cube sits on top of base cubes · Apex cube centre is within tolerance of target marker · Apex-target overlap ≥ 20%

Each task ships 50 evaluation scenes (dataset/robowits/eval_dataset_50/<NN>.json), which is where placement variation comes from: same predicate, different object poses.

Demo coverage

23 of the 30 tasks have a clip on the project page, downloaded by make demos and played inline. The seven without one are exactly the seven that use non-rigid physics — dough (MPM), sand, water and marbles (SPH/PBD):

round_dough_sheet · separate_marbles_and_sand · ball_into_jar · seal_colander · stabilize_bottle · water_into_mug · differentiate_cubes

They are also the ones we have not run, and the correlation is worth keeping in mind when triaging: the deformable tasks are the least evidenced part of the suite.

The project page also publishes two mutation clips for several seed tasks (same predicate, added distractors, a replaced tool). We do not mirror those; they are one click away from the task's own row.

What we run it on

Pinned to the commit our image builds from, 9cc30ae (the 2026-06-04 code release), and measured on an RTX 4090 laptop:
Image nvidia/cuda 12.9 + RoboWits from its own uv.lock, ~20 GB
Physics CPU backend — Genesis' GPU backend settles contacts differently run to run
Determinism Bit-exact across processes on the CPU backend
Throughput ~700 control steps/s headless, ~30 with three cameras rendering
Scene build ~60 s per process

27 of 30 tasks need BlenderKit assets

Their meshes are paid assets that may not be redistributed, so they are not in the image: the host owner downloads them with a personal API key. Only 06 dominos, 07 stand pages and 14 stack cubes run from the Apache-2.0 bundle alone — which is why 14 stack cubes is the task we integrated first.