RoboCasa365

seq init robocasa gives the stable version, seq init robocasa@beta the beta, and seq init robocasa@v<N> any version below that is not withdrawn. A benchmark registered from this template names the version it runs in seq benchmark ls; seq benchmark upgrade moves it to the newer stable one.

1. v72026-10-13beta

   robosuite and RoboCasa, the source both images install, come to you: seq init downloads each into third_party/ from its official package, the very package the images install.

   Needs sequence-ai 0.34.17 or later; 0.34.20 or later for seq init to download the source.

   • upstream.toml names robosuite at 5ce6643f3092639d08f7b0f90ed1c6a84f50552c and RoboCasa at 456174f62b89b8fca99eaaf33949c29fec9cfc2a, the commits v6 pins, each as one package, by its sha256 and the sha256 of the tree it unpacks to (third_party/SOURCES.json). seq init downloads both into third_party/ and checks them; seq init --no-source skips them.
   • The images install robosuite and RoboCasa from those same packages, frozen with this version; v6 installed a copy of the same commits, packed when it was frozen. The files are the same.
   • seq deploy leaves third_party/robosuite and third_party/robocasa out of what it uploads while they are as published, and stops while one was changed: the images install only the published packages, so a change could not run.
   • The kitchen assets are not in RoboCasa's package: the six archives its download script fetches stay with this version's frozen files, as in v6 (NOTICE).
   • requirements-benchmark.lock also lists the platform's own packages the GPU environment holds (fastapi, uvicorn, pydantic and what they need), at the versions the platform installs - the environment v6's image held, now locked whole.
   • Everything else - the Python and system packages, the base, the tasks, scenes and success - is v6's.
   Use v7

   robocasa v7shell

   # start a benchmark from this version, then register it
   seq init robocasa@v7 my-benchmark
   seq deploy my-benchmark/benchmark.py
   # its CPU version, on CPU machines
   seq deploy my-benchmark/benchmark_cpu.py
   
   # or move a benchmark registered from this template to it
   seq benchmark upgrade my-benchmark --to v7

2. v62026-10-13stable

   Both images install only from this version's frozen files, which the platform keeps: v5's packages, robosuite and RoboCasa at their commits and the kitchen assets, the same on any day.

   Needs sequence-ai 0.34.17 or later.

   • benchmark.py and benchmark_cpu.py start from debian:12-slim by digest and install from /seq/frozen (.template_files("robocasa", version=6)): the system packages as .deb files; every Python package with no index, each file's sha256 checked against requirements-benchmark.lock, requirements-benchmark-classic.lock and requirements-benchmark_cpu.lock; robosuite and RoboCasa from their sources at the commits in robocasa_tasks.py; and RoboCasa's six kitchen asset archives, unpacked where its download script puts them. frozen.toml says what is frozen.
   • The Python packages are the ones v5's images hold, read out of them (pins-*.txt): torch 2.7.1 among them, as in v4 and v5. TORCH_CPU is gone - nothing installs torch ahead of the locked packages.
   • To add a package or change a version, add your own step after the template's (.uv_pip_install(...)); it installs from PyPI as any image does.
   • NOTICE is v5's. Tasks, splits, seeds, success and the step limit are as in v5.
   Use v6

   robocasa v6shell

   # start a benchmark from this version, then register it
   seq init robocasa@v6 my-benchmark
   seq deploy my-benchmark/benchmark.py
   # its CPU version, on CPU machines
   seq deploy my-benchmark/benchmark_cpu.py
   
   # or move a benchmark registered from this template to it
   seq benchmark upgrade my-benchmark --to v6

   In the console, New benchmark registers this version: the console registers the stable one.

3. v52026-10-12

   NOTICE says where each set of RoboCasa's kitchen assets comes from and under which license. The benchmark is v4's, and its images install what v4's did.

   Needs sequence-ai 0.33.22 or later.

   • fixtures_lw and objs_lw: fixtures and objects by Lightwheel, from NVIDIA's dataset PhysicalAI-Robotics-Manipulation-Objects-Kitchen-MJCF, under CC BY 4.0.
   • objs_objaverse: objects from Objaverse 1.0, which is released under ODC-By 1.0. Each object keeps the license its creator chose, some of them non-commercial (CC BY-NC) or share-alike (CC BY-SA), and RoboCasa's archive does not list each one's.
   • objs_aigen (objects generated with Luma AI) and the textures tex and tex_generative (the latter generated with AI tools): RoboCasa releases them under CC BY 4.0.
   • robocasa_tasks.py pins the CPU build of torch both images install first (TORCH_CPU = "2.14.1"). lerobot 0.3.3 needs a torch before 2.8, so the install after it replaces it with torch 2.7.1, as in v4.
   Use v5

   robocasa v5shell

   # start a benchmark from this version, then register it
   seq init robocasa@v5 my-benchmark
   seq deploy my-benchmark/benchmark.py
   # its CPU version, on CPU machines
   seq deploy my-benchmark/benchmark_cpu.py
   
   # or move a benchmark registered from this template to it
   seq benchmark upgrade my-benchmark --to v5

4. v42026-10-09

   Both images install their Python dependencies as PyPI had them at a fixed time, so a rebuild on any day installs exactly the same packages.

   Needs sequence-ai 0.33.22 or later.

   • benchmark.py (robocasa): the GPU side takes its packages as of 2026-10-07 03:30 UTC (UV_EXCLUDE_NEWER), and RoboCasa's own environment in the same image as of the same moment (pip 26.2.1's --uploaded-prior-to).
   • benchmark_cpu.py (robocasa-cpu): its packages as of 2026-10-06 22:49 UTC.
   • Unpinned, a rebuild on 2026-10-09 took pyarrow 26.0.0 into both versions, pydantic 2.14.0 and a newer sequence-ai and sequence-base into the GPU version's RoboCasa environment, and newer absl-py, imageio and platformdirs.
   • To take newer versions, move _PYPI_AS_OF forward in each file and check the benchmark again.
   Use v4

   robocasa v4shell

   # start a benchmark from this version, then register it
   seq init robocasa@v4 my-benchmark
   seq deploy my-benchmark/benchmark.py
   # its CPU version, on CPU machines
   seq deploy my-benchmark/benchmark_cpu.py
   
   # or move a benchmark registered from this template to it
   seq benchmark upgrade my-benchmark --to v4

5. v32026-10-09

   benchmark.py registers from a machine with nothing but sequence-ai installed: it no longer needs numpy where you run seq deploy.

   Needs sequence-ai 0.33.22 or later.

   • seq validate and seq deploy import benchmark.py on your own machine to read its declaration. v2 imported numpy at the top, so with only sequence-ai installed they stopped with No module named 'numpy'; numpy is now used only on the platform, where the image installs it.
   • The benchmarks, their images and the adapter robocasa-openpi are v2's.
   Use v3

   robocasa v3shell

   # start a benchmark from this version, then register it
   seq init robocasa@v3 my-benchmark
   seq deploy my-benchmark/benchmark.py
   # its CPU version, on CPU machines
   seq deploy my-benchmark/benchmark_cpu.py
   
   # or move a benchmark registered from this template to it
   seq benchmark upgrade my-benchmark --to v3

6. v22026-10-08

   The same benchmark as v1, published again without the notes v1's files carried for our own records.

   Needs sequence-ai 0.33.22 or later; to register benchmark.py, numpy on your machine too.

   • benchmark.py, benchmark_cpu.py, the adapter robocasa-openpi, the tasks and the episodes are v1's. Only comments and the README differ, without those notes.
   Use v2

   robocasa v2shell

   # start a benchmark from this version, then register it
   seq init robocasa@v2 my-benchmark
   seq deploy my-benchmark/benchmark.py
   # its CPU version, on CPU machines
   seq deploy my-benchmark/benchmark_cpu.py
   
   # or move a benchmark registered from this template to it
   seq benchmark upgrade my-benchmark --to v2

7. v12026-10-07withdrawn

   The first version: RoboCasa365's 18 atomic tasks as a benchmark, in its pretrain kitchens by default, with robocasa on one GPU (MuJoCo Warp), robocasa-cpu on CPU machines (RoboCasa itself) and the adapter robocasa-openpi.

   Withdrawn on 2026-10-08: its files carried notes written for our own records. v2 is the same template without them.

   • seq deploy benchmark.py registers robocasa: an instance runs its environments as MuJoCo Warp worlds on one L4, 16 by default. The image exports the pretrain kitchens' scenes from RoboCasa, and RoboCasa's own renderer draws the cameras.
   • seq deploy benchmark_cpu.py registers robocasa-cpu: RoboCasa365 unchanged, one environment per core, 8 cores and 64 GB an instance.
   • The tasks: atomic_seen's 18, 50 episodes each (5 in each of 10 kitchens), each capped at RoboCasa365's horizon (450-1,050 steps at 20 Hz) and scored by RoboCasa's own success rule. --param benchmark.split=target plays the target split's kitchens on robocasa-cpu.
   • seq deploy pi05_robocasa_adapter.py registers the adapter robocasa-openpi: the left, right and wrist cameras padded to 224x224, the 16-D state and the instruction, as RoboCasa's openpi client gives a pi0.5; its 12-D actions at 20 Hz go to the simulator unchanged. Recordings film the left agent view at 20 fps.
   • pi05-robocasa scored 24 of 54 episodes (44.4%) on robocasa-cpu, 3 seeds a task, measured on the platform; RoboCasa365 publishes 39.6% for that checkpoint.
   • The template's own code is Apache-2.0. NOTICE carries the notices of RoboCasa365 (code MIT; kitchens, objects and textures CC BY 4.0) and robosuite (MIT).