HexaShell An agent harness for robot manipulation: RoboDojo's two-arm table tasks, RoboCasa365, LIBERO-PRO, RoboTwin 2.0, and a real robot through URAI. HexaShell's code is its own, but it rests on other projects' files, measurements and ideas: RoboDojo (MIT License, Copyright (c) 2026 RoboDojo Contributors) - hexashell/assets/X5A.urdf is a copy of the ARX X5 robot description in RoboDojo's assets. - hexashell/assets/robot_model.npz holds triangles derived from RoboDojo's ARX X5 meshes (made by bench/build_robot_points.py). - xpolicylab/HexaShell follows the layout of RoboDojo's XPolicyLab policy template. - scripts/patches/robodojo-layout-ids.patch changes RoboDojo's env/seed_manager/seed_manager.py so that ROBODOJO_LAYOUT_IDS selects the layouts of an evaluation (used by scripts/run_one.sh and scripts/batch.sh; RoboDojo's own evaluation runs without it). RoboProbe (Apache License 2.0) - prompts/ladders/*.md are RoboDojo's task descriptions and scoring tables, copied unchanged from RoboProbe; see prompts/ladders/README.md and prompts/ladders/LICENSE. PyRUA-Lean (Apache License 2.0, Copyright 2026 PyRUA-Lean Contributors) Several facts about the simulator and several design ideas were first worked out there: - simulator facts: the arm-base poses (re-checked against our own recordings with bench/fit_bases.py), the joint ranges the simulator lets the arms reach, the grasp-point and fingertip offsets along link6, the squeezed-gripper width fit, the wrist-camera mount, the head-camera pose and both focal lengths; - motion constants: at most 0.15 rad per joint per action, up to 8 actions to settle, arrival within 1 cm / 5 degrees, near misses accepted within 2 cm / 8 degrees, lifting before swinging home; - ideas: the model works the robot by writing Python cells against a robot object; objects are measured in 3-D from several RGB cameras as the shape that agrees with all their outlines, with the robot's own pixels left out and measured heights remembered; a bowl's reading from its rim, kept only where it explains the views clearly better; with one camera, the height read from how far the outline's upright sides reach out on the table plane; - prompt advice adopted after comparing our prompt with a port of PyRUA-Lean's: coming down onto an upright object from above for a level side grasp, measuring every object before the arms move, judging a pour by what is left when the container turns back up, homing before a benchmark-driven arm's turn, that a bumped container moves, switching arms or reaching with a tilted or level hand when one arm cannot reach, that the way a task is done can be checked, and measuring a low object from the top grasp's approach pose; skills/robodojo/hang_mugs.md's level body grasp and level slide onto the peg were seen in a run with that port. HexaShell-Preview (HexaShell 1.0; this repository's earlier version, branch v1.0: https://github.com/HexaFuture/HexaShell/tree/v1.0) - ideas carried over: dry-run reachability checks, two-view triangulation with a ray-gap check, ray-plane intersection at a given height, a mesh-based robot mask, support-height detection, and the hover advice for re-measuring from above. RoboCasa365 and robosuite (MIT License) - hexashell/benches/robocasa runs RoboCasa's kitchens, tasks and success checks unchanged, through robosuite's official controllers; the 50 tasks, their sets and step limits are RoboCasa365's benchmark split. Both are included only as git submodules (third_party/robocasa, third_party/robosuite) at upstream commits, with the changes to RoboCasa listed in THIRD_PARTY_VERSIONS.md applied by scripts/patch_third_party.py; RoboCasa's assets are not included. LIBERO and LIBERO-PRO (MIT License) - hexashell/benches/libero runs LIBERO-PRO's suites through LIBERO's own environments, controller and goal checks (the rpent-libero and rpent-liberopro packages); neither is part of this repository; - the step limits of the test (OFFICIAL_STEPS of hexashell/benches/libero/protocol.py) and its 10 settling steps are those of LIBERO-PRO's own evaluation. RoboTwin 2.0 (MIT License, Copyright (c) 2025 Tianxing Chen) - hexashell/assets/robotwin_aloha_arms.urdf holds the joints of the two front arms of RoboTwin's aloha-agilex robot, copied from its assets (embodiments/aloha-agilex/urdf/ arx5_description_isaac.urdf; meshes, inertia and the rest of the robot left out); - hexashell/benches/robotwin runs RoboTwin's tasks, controller, ray-traced cameras and success checks unchanged, through the rlinf-robotwin-runtime package (MIT); neither is part of this repository; - STEP_LIMITS of hexashell/benches/robotwin/protocol.py are RoboTwin's step limits (_eval_step_limit.yml), and the test takes its seeds and instructions as RoboTwin's evaluation does (scripts/eval_policy.sh: the first seeds from 100000 its expert solves; for demo_randomized, instructions from the tasks' 'unseen' description templates). RPent (Apache License 2.0) - hexashell/benches/robotwin/eval_seeds.json is a copy of RPent's RoboTwin evaluation list (robots/robotwin/eval/demo_randomized.json at commit 886b3b2: 50 tasks, five checked seeds each, with their instructions), which gives the task list and a check of the seed scan. URAI (Apache License 2.0, Copyright 2026 Hexafuture Inc. and the URAI authors, https://github.com/HexaFuture/URAI) - hexashell/real/urai/capsules.py is a copy of the ray-capsule intersection and its pixel windows from URAI's urai/robot/scene.py (Capsule, _capsule_hit, _capsule_window, _camera_rays, _pixel_rays, _window), with public names and without the capsules' colour; - prompts/urai/*.md are the English real-robot prompts of URAI's HexaShell harness, copied unchanged, and the sentences of hexashell/real/urai/texts.py are that harness's English ones; - hexashell/real/urai follows that harness and talks to URAI only over its HTTP interface; URAI itself is not part of this repository. SAM3 (Meta) runs as a separate local service (services/sam3_server.py); its code and weights are not part of this repository.