Cortex splits long-horizon manipulation at a subtask interface: aligned planning and execution reach 76.8 on a real-world chemistry-style task
Synopsis
InternRobotics released Cortex, also called InternVLA-M1.5, a bidirectionally aligned embodied agent framework for long-horizon manipulation in which a high-level vision-language model acts as a cognitive orchestrator that tracks progress, updates semantic memory and emits executable subtasks, while a low-level vision-language-action model handles reactive physical execution. The framework standardizes manipulation behavior into 32 skill primitives and ships more than 4K hours of long-horizon subtask data. Cortex scores 95.5, 86.8 and 76.8 on LIBERO-long, RoboTwin 2.0 and real-world tasks, where the same comparison table lists pi05 at 22.1 on the real-world setting.
Interpretation
Long-horizon manipulation is split at a shared subtask interface: the high-level planner only judges how far the task has progressed, updates memory and names the next executable subtask, while the low-level policy only carries that step out. The prior route compressed perception through action into one end-to-end policy, so step ordering, completion checks and discrimination between similar stages all fell to a single model. On a real-world multi-stage chemistry-style task Cortex scores 76.8, against 22.1 for pi05 in the same comparison table; on LIBERO-long and RoboTwin 2.0, Cortex scores 95.5 and 86.8 against 92.4 and 82.7 for pi05.
The high-level planner is no longer trained only by sampling uniformly from trajectories; training emphasizes the transition regions where the planner must verify that the previous step finished and rewrite its memory. The prior training sampled only uniformly from trajectories, so judgments at stage switches received no separate emphasis. At step level Cortex reaches an average total of 8.318, against 6.925 for Gemini-3.1-Pro, 6.739 for Qwen3-VL-8B-Instruct and 6.268 for GPT-5; at episode level Cortex reaches 7.810, against 7.231 for GPT-5, 6.860 for Gemini-3.1-Pro and 6.292 for Qwen3-VL-8B-Instruct.
Manipulation behavior is organized into a standardized long-horizon subtask space of 32 canonical skill primitives, and each subtask carries grounding for space, number and object attributes. The prior subtask space was not unified into a fixed set of skill primitives and did not attach spatial, numerical and object-attribute grounding to subtasks. Evaluation is organized around spatial grounding, long-horizon logical consistency and object counting, and both the step-level and episode-level scores are reported separately on those three task types.
Perspective
The demonstrated scope is two long-horizon simulation suites, LIBERO-long and RoboTwin 2.0, plus a zero-shot real-world multi-stage chemistry-style task; evaluation is organized around spatial grounding, long-horizon logical consistency and object counting, with step-level and episode-level scores reported separately.
The real-world experiment is described as a zero-shot multi-stage chemistry-style task, and its sample size and number of repetitions are not given; step-level evaluation uses ground-truth memory, while episode-level evaluation measures semantic drift when the planner reads its own previous memory, so the two measure different capabilities.
