Mosi-AI / RoboICL

RoboICL: Embodied In-Context Learning with GPT-6 Astra 
 Multimodal Models as Few-Shot Robot Learners 
 RoboICL uses interaction memory alone on Open and one demonstration on the other categories. 
## News
- **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.
## Overview
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.
## Installation
### Prerequisites
- 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.
### Clone the repository
 [code] 
For an existing clone:
 [code] 
The exact supported commits of RoboDojo, XPolicyLab, IsaacLab, and cuRobo are
recorded in `configs/upstream.lock.json` and verified before launch.
### Configure the two Python environments
Point RoboICL at the interpreters that contain your simulator and policy
dependencies:
 [code] 
Install the direct RoboICL dependencies into those environments after the
Isaac Sim environment itself is available:
 [code] 
If the simulator environment provides a newer `libstdc++` than the host:
 [code] 
## Data preparation
### RoboDojo assets
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.
 [code] 
### One-shot references
The paper benchmark contains exactly 8 Open tasks evaluated zero-shot and 22
non-Open tasks supplied with o