---
license: mit
task_categories: [robotics]
tags: [lerobot, so100, so101, maniskill, simulation, pick-and-place, language-conditioned, domain-randomization, ci-mse]
pretty_name: SO-100 "put the {color} cube in the bowl" (ManiSkill sim) — v2
---
> **⚠️ ERRATUM (2026-09-13): task labels in this dataset are misaligned with the episodes.**
> The ManiSkill→LeRobot converter stored trajectories in lexicographic key order (`traj_0, traj_1, traj_10, …`) while the
> per-episode language labels were assigned in numeric order, so about two thirds of episodes carry the wrong cube colour
> in `task` / `task_index` / `meta/episode_targets.json`. Actions, states and videos are correct and internally consistent;
> only the language label is permuted. Corrected per-episode targets are in `meta/episode_targets_fixed.json`
> (dataset episode *k* holds raw trajectory `lex[k]`, `lex = sorted(range(N), key=lambda i: f"traj_{i}")`).
> A relabelled release will replace the `task` columns; until then do not use the `task` column as-is for training or
> evaluation. Results in `Kavin60606/cimse-so100-experiment-v2` that depend on language grounding are affected (see its README).


# SO-100 pick-cube-into-bowl (ManiSkill3, domain-randomized) — v2 train

Language-conditioned pick-and-place on a simulated SO-100: three coloured cubes (red/green/blue) and a bowl at random poses,
instruction `put the {color} cube in the bowl`. Built for an offline-metric ↔ closed-loop-success study (CI-MSE) with a
MolmoAct2 checkpoint zoo. Every episode succeeds (target cube inside the bowl, released, arm at rest).

| | |
|---|---|
| env | `PickCubeBowlSO100-v1` (custom ManiSkill3 env, code in `fd-studio/eval/sim_so100_v2/sim/`), 30 Hz, max 450 steps |
| cameras | `observation.images.cam0` front/top-down, `observation.images.cam1` side, 224×224, per-episode pose ±2 cm / ±3°, fov ±2.5° |
| state / action | 6-D joint positions / absolute joint targets, **degrees** (pan, lift, elbow, wrist_flex, wrist_roll, gripper; gripper 0 open, −46 closed) |
| DR | lighting (ambient, directional), cube size 1.25–1.45 cm, colour jitter, bowl grey jitter, camera pose |
| tasks | 3 (`meta/tasks.parquet`); per-episode task = target colour; `meta/episode_targets.json` |
| stats | q01/q99 quantiles included (MolmoAct2 normalization) |
| filter | episodes > 450 steps dropped (3–5 %) |

## Demonstrator sources
| source | how | oracle SR |
|---|---|---|
| **S0 clean** | screw→RRT motion planner, ±1 cm waypoint jitter, dz×yaw release candidates | 0.56 |
| **S1 noisy** | + joint noise σ 2.5°/step, pauses 0.2–0.5 s, 20 % first-miss + regrasp, 10 % random carry waypoint | 0.40 |
| **S2 strategy-varied** | planner speed 0.4–1.6×, carry height 3–8 cm, grasp yaw +90° (30 %), lateral approach then slide (50 %) | 0.44 |

**Train split**: S0 1200 + S1 900 + S2 900 episodes (seeds 1e6+, 2e6+, 4e6+), ~3000 before the length filter. Sources are mixed; `meta/episode_targets.json` gives target colour per episode, source order is S0 then S1 then S2 by episode index.

Related: `Kavin60606/cimse-so100-experiment-v2` (raw h5, videos, proof grids, verify reports, eval seeds), v1 study `Kavin60606/cimse-so100-experiment`.