Mosi-AI/RoboICL

Multimodal Models as Few-Shot Robot Learners

RoboICL uses interaction memory alone on Open and one demonstration on the other categories.

• 2026-10-07 For reproducible rollouts, we recommend the official API or a stable third-party relay, such as DMXAPI's gpt-6-astra-ssvip, rather than an unstable coding-plan endpoint.
• 2026-09-29 Paper released on arXiv, together with the code and 800+ interactive evaluation rollouts.
• 2026-09-19 Research Preview launched.

RoboICL is an in-context robot-control framework that adapts a frozen general-purpose vision-language model from demonstrations and its own execution history. It requires no robot-specific parameter update and no learned vision-language-action model.

RoboICL separates context into two complementary sources:

• Demonstration context provides recorded robot trajectories as executable few-shot examples.
• Interaction memory retains the model's own actions, controller feedback, and resulting observations during the current rollout.
• A shared interaction grammar presents both sources as observation -> action -> receipt -> observation records.
• Bounded anchored memory preserves selected early and intermediate interactions while retaining the latest result for immediate correction.

The released runner connects GPT-6 Astra to RoboDojo through one validated action interface. At each request, the model observes a triptych of the left wrist, head, and right wrist cameras together with robot proprioception, then predicts a bounded sequence of dual-arm Cartesian actions for closed-loop execution.

• Linux with an NVIDIA GPU and a working Vulkan/CUDA stack
• Git and Git LFS
• an Isaac Sim 5.1-compatible simulator environment
• a separate Python environment for the multimodal policy process

The simulator and policy processes intentionally use separate interpreters because their websockets and simulator dependency constraints differ.

git clone --recurse-submodules https://github.com/Mosi-AI/RoboICL.git
cd RoboICL

For an existing clone:

git submodule update --init --recursive

The exact supported commits of RoboDojo, XPolicyLab, IsaacLab, and cuRobo are recorded in configs/upstream.lock.json and verified before launch.

Point RoboICL at the interpreters that contain your simulator and policy dependencies:

export ROBOICL_SIM_PYTHON=/path/to/sim/bin/python
export ROBOICL_POLICY_PYTHON=/path/to/policy/bin/python
export ROBOICL_DATA_ROOT="$PWD/data"
export ROBOICL_RESULTS_ROOT="$PWD/results"
source setup/paths.sh

Install the direct RoboICL dependencies into those environments after the Isaac Sim environment itself is available:

"$ROBOICL_POLICY_PYTHON" -m pip install -r setup/requirements-policy.txt
"$ROBOICL_SIM_PYTHON" -m pip install \
  -r setup/requirements-sim.txt \
  -c setup/constraints-sim.txt

If the simulator environment provides a newer libstdc++ than the host:

SIM_PREFIX="$(dirname -- "$(dirname -- "$ROBOICL_SIM_PYTHON")")"
export ROBOICL_SIM_LD_PRELOAD="$SIM_PREFIX/lib/libstdc++.so.6"

Fetch the complete frozen RoboDojo asset tree into an external Git/LFS cache. The script checks out the revision in configs/data.lock.json, verifies the download, and links data/Assets to the cache.

"$ROBOICL_POLICY_PYTHON" setup/fetch_assets.py \
  --data-root "$ROBOICL_DATA_ROOT" \
  --cache /path/to/robodojo-data-cache

The paper benchmark contains exactly 8 Open tasks evaluated zero-shot and 22 non-Open tasks supplied with one demonstration. The complete source-episode identity, HDF5 SHA256, prediction horizon, and all 12 selected action-window starts are archived in configs/references/one_shot_j12_b12.json.

Fetch and deterministically build any archived one-shot reference, for example:

"$ROBOICL_POLICY_PYTHON" setup/fetch_reference.py deposit_coin \
  --data-root "$ROBOICL_DATA_ROOT"

"$ROBOICL_POLICY_PYTHON" setup/fetch_reference.py fill_pen_holder \
  --data-root "$ROBOICL_DATA_ROOT"

The source is checked by SHA256, and the builder uses the exact archived starts rather than silently re-running a heuristic. The generated bundle contains 12 non-overlapping action blocks and their starting/result observations.

TRAIN and LIVE use different selectors. TRAIN uses the exact action-window starts recorded in the setting manifest. LIVE does not sample video frames: eleven uniformly spaced time anchors retain the first complete executed interaction that crosses each anchor, while the twelfth slot retains the latest complete interaction.

To build a reference for another task from a local HDF5 episode:

"$ROBOICL_POLICY_PYTHON" -m roboicl.reference build \
  --task "$TASK" \
  --source "$ROBOICL_DATA_ROOT/runtime-data/$TASK/train/episode_0000000.hdf5" \
  --source-id "runtime-data/$TASK/train/episode_0000000.hdf5" \
  --source-sha256 "$SOURCE_SHA256" \
  --output "$ROBOICL_DATA_ROOT/runtime-data/$TASK/reference_1shot_train12_endpoint_h$H" \
  --horizon "$H" \
  --chunks 12 \
  --selected-starts "$START_0,$START_1,...,$START_11"

The release provides two main protocols:

Both use the task-adaptive action horizons in configs/adaptive_horizons.json.

"$ROBOICL_POLICY_PYTHON" -m roboicl.run deposit_coin \
  --shots 0 \
  --seed 0 \
  --layout 0 \
  --profile configs/protocols/zero_shot_b25.json \
  --dry-run

This validates the task, layout, assets, profile, interpreters, and pinned submodules without starting the simulator or calling the model API.

"$ROBOICL_POLICY_PYTHON" -m roboicl.run deposit_coin \
  --shots 0 \
  --seed 0 \
  --layout 0 \
  --gpu 0 \
  --profile configs/protocols/zero_shot_b25.json \
  --capture-only

--capture-only launches the scene and verifies the three RGB streams without making a model request or executing an action.

export ASTRA_API_KEY=...

"$ROBOICL_POLICY_PYTHON" -m roboicl.run deposit_coin \
  --shots 0 \
  --seed 0 \
  --layout 0 \
  --gpu 0 \
  --profile configs/protocols/zero_shot_b25.json

export ASTRA_API_KEY=...

"$ROBOICL_POLICY_PYTHON" -m roboicl.run deposit_coin \
  --shots 1 \
  --seed 0 \
  --layout 0 \
  --gpu 0 \
  --profile configs/protocols/one_shot_j12_b12.json

Replace the task, seed, and layout as needed. Use --reference /path/to/bundle for a one-shot task without a published reference path. A process exit code of zero means evaluation completed; task success and progress are reported by the native RoboDojo evaluator in the generated result.json.

roboicl/
├── run.py                 preflight and process orchestration
├── robodojo.py            single-layout RoboDojo adapter
├── rollout.py             observation/action execution loop
├── policy_server.py       XPolicyLab WebSocket policy entrypoint
├── reference.py           reference builder and verifier
└── policy/                prompting, memory, validation, and transport
configs/
├── protocols/             zero-shot and one-shot protocol profiles
├── tasks/                 official documentation for 42 tasks
├── adaptive_horizons.json task-specific action horizons
├── data.lock.json         frozen asset revision
├── references/             setting-level source episodes and exact TRAIN windows
└── upstream.lock.json     pinned submodule commits
setup/
├── fetch_assets.py
├── fetch_reference.py
├── requirements-policy.txt
├── requirements-sim.txt
└── paths.sh
third_party/
├── RoboDojo/
├── XPolicyLab/
├── IsaacLab/
└── curobo/

@article{liu2026roboicl,
  title={RoboICL: Embodied In-Context Learning with GPT-6 Astra},
  author={Liu, Fangcheng and Shen, Yeqing and Cheng, Anda and Mi, Weishi and Tang, Chao and Liu, Chenyuan and Xiang, Yushun and Li, Tingguang and Li, Yong-Lu and Tang, Yehui},
  journal={arXiv preprint arXiv:2609.34261},
  year={2026}
}

RoboICL's project-owned code is released under the MIT License. Third-party submodules and benchmark assets retain their original licenses and terms.