← All model news

ZTE EmbodiedMind Moves GRPO Advantage Calculation to Action Prefixes: A Three-Stage Training Recipe for an 8B Embodied Model

According to a ZTE article published by Machine Heart on November 6, 2026 in its "release" format, the EmbodiedMind paper proposes a three-stage training paradigm—RSFT, IR-GRPO, and Trie-GRPO—claiming that the 8B model averages 70.02% across 18 embodied benchmarks and improves long-horizon planning accuracy by 30.31% over current best methods. It also claims Nebula-EmbodiedBrain-180B tied for first in China on the September SuperCLUE embodied brain leaderboard. The original article does not mention open weights, code, or API access.

Published Updated

What changed

On November 6, 2026, Machine Heart published a ZTE embodied-model article in its "release" format. According to the article, the ZTE team's paper, "EmbodiedMind: Adaptive Data Curation and Prefix-Tree Reinforcement Learning for Efficient Embodied Intelligence," proposes a three-stage training paradigm. First, RSFT stratifies samples by running multiple inferences at high temperature: samples that are all correct are labeled easy, partially correct ones uncertain, and all-wrong ones hard; the training set keeps only uncertain samples plus a proportion of easy samples. Next, IR-GRPO maintains an independent FIFO queue for each task type, uses an exponential moving average to smooth each task's effective sample ratio, and dynamically adjusts the next round's sampling quota. Finally, Trie-GRPO inserts all sampled trajectories in parallel into an action prefix tree, where trajectories sharing the same history prefix form sibling subgroups at branching points; relative advantages are computed only among sibling nodes and propagated back along tree paths via discounted returns, with near-zero-advantage samples filtered out to control exploration cost. The article claims that EmbodiedMind-8B uses only 522,000 training samples and averages 70.02% across 18 benchmarks covering general multimodal understanding, 2D spatial reasoning, 3D spatial measurement, and task planning, with long-horizon planning accuracy 30.31% higher than current best methods; the paper has been submitted to ICRA 2027. It also gives specific comparisons: 66.67 on VLM-PlanSim-99 with up to 25 steps, above Cosmos3-Nano; 60.45% success on ALFWorld-thor versus 46.27% for SEEA-R1; and on 50 desktop scenarios built on AgiBot A2 plus 160 task-execution monitoring videos, 62.00% on desktop tidying planning (versus 40.00% for Qwen3-VL-8B) and 55.31% task-execution monitoring accuracy (versus 46.88%). The same article also claims that Nebula-EmbodiedBrain-180B scored 77.48 on the SuperCLUE September 2026 EmbodiedCLUE-VLA embodied brain leaderboard, tying with Qwen3.8-Max-0902 for first in China, with 97.92 on the embodied safety sub-item.

What it means for developers

First, separate "what can be integrated" from "what can be borrowed." The original article does not mention open weights, a code repository, or a hosted interface for EmbodiedMind-8B or Nebula-EmbodiedBrain-180B, so these two names currently do not correspond to anything you can call directly or deploy locally; what can be analyzed here is the training method and evaluation methodology. If you are doing reinforcement-learning post-training for embodied or VLA models, whether "groups" are defined at the whole-trajectory level or the action-prefix level directly changes the gradient composition of failed trajectories: per the article's explanation, trajectory-level advantages penalize correct intermediate actions in a failed trajectory along with everything else, while the prefix tree confines comparison to sibling nodes and filters near-zero-advantage samples to keep exploration cost manageable. This change does not depend on a particular base model and in principle can be ported into your own GRPO-style implementation, at the cost of requiring environments capable of branching sampling. For developers focused on evaluation, another usable lead is the third-party leaderboard: the September SuperCLUE methodology places 180B-class models in the same tier as 2.4T-parameter models, and the original article explicitly states that international reference models do not participate in the domestic ranking, yet its listed reference scores (GPT-6 Astra 86.75, Gemini-3.8-Flash 82.12) are higher than 77.48—when reading the leaderboard, separate ranking scope from absolute scores. Overall, this is a methodological advance, not a model release you can integrate directly.

Integration and evaluation

The following are conditional suggestions, not reported facts. Suggestion one: whether to move GRPO's advantage grouping from whole trajectories to action prefixes. The change reported in the article is Trie-GRPO: standard GRPO applies trajectory-level advantage uniformly to all tokens, so a trajectory that ultimately fails also penalizes its correct intermediate decisions; the prefix tree lets trajectories sharing the same history prefix form sibling subgroups at branching points and normalizes relative advantage only among sibling nodes. The choice to make is whether RL post-training for multi-step tasks groups by trajectory or by replayable action prefixes. The prerequisites for changing this choice are environment capability and task length: prefix grouping only makes sense when the simulator or real environment can reset from intermediate states and branch from the same prefix, and when enough sibling samples can be collected under the same prefix for normalization; when tasks are mostly one or two steps, or the environment cannot branch from the middle, trajectory-level grouping is easier to implement. A measurable criterion: under the same sampling budget, run both RL variants on your own long-horizon evaluation set of 10+ steps and compare final success rate and step-wise accuracy; the article's 30.31% is the authors' improvement figure and should not be taken directly as your expected gain. Suggestion two: whether to add an uncertainty filter before training. The article reports RSFT: multiple high-temperature inferences split samples into all-correct, partially correct, and all-wrong categories; the training set keeps only partially correct samples plus a proportion of easy samples, temporarily excluding all-wrong samples. The choice to make is whether to add this filter before SFT and how many easy samples to keep. The prerequisite is a reliable scoring signal—stratification depends entirely on judging right/wrong, so noisy scoring distorts the classification itself, and multiple high-temperature inferences cost extra inference budget up front. A measurable criterion: record the number of effective samples after filtering and the training epochs, compare against SFT on the full corpus across capability evaluations, and watch for signs of distribution collapse such as a sharp drop in output diversity early on. Suggestion three: whether to add a task-level quota layer to dynamic sampling. The article reports IR-GRPO: an independent FIFO queue per task type, an exponential moving average to smooth each task's effective sample ratio, and dynamic adjustment of the next round's sampling quota. The choice to make is whether to add quota control by task label on top of DAPO-style dynamic sampling. The prerequisites are reliable task-type labels and naturally imbalanced task difficulty; when tasks are highly homogeneous, this layer yields limited benefit. A measurable criterion: record per-task-type scores round by round, check for a seesaw where one category rises as another falls, and compare the worst task type's score floor with and without quotas.

Limits and open questions

This article was published by Machine Heart in its "release" format; the methods and numbers come from the ZTE team's paper and the original article's statements. This article has not independently reproduced them or verified any of them on real robots. The paper's status is "submitted to ICRA 2027," not accepted or peer-reviewed, and the benchmark scores are author-reported; the original article also does not state whether 30.31% is a relative improvement or an absolute percentage-point difference, so that uncertainty should be kept in mind when reading the number. The real-robot conclusions rest on the A2-Failure dataset of 50 desktop scenarios and 160 monitoring videos—a limited scale—and transfer to other environments is unknown. The SuperCLUE leaderboard reflects September 2026, earlier than the article's publication date, and the original article explicitly states that international models are reference-only and do not participate in the domestic ranking, so leaderboard scores should not be treated as durable conclusions across time. The original article does not mention open weights, code, or API access; as of a check of the KimiSeek live catalog on November 12, 2026, neither EmbodiedMind-8B nor Nebula-EmbodiedBrain-180B is listed, and if a hosted interface appears later, its availability is subject to the provider's terms.

Three minutes to your first working call

Create an account and an API key. Calls use your account balance. 30% off your first top-up. Top up ¥500, pay ¥350, receive ¥500 in credit. Top up ¥1000, pay ¥850, receive ¥1000 in credit You can still start with a ¥10 trial top-up.

Sources and editorial process

让机器人学会「一口气干完一件活」,中兴 EmbodiedMind 重构具身大脑训练范式

AI assisted the analysis and translation of the cited sources, followed by automated quality checks. Recommendations are not test results. Consult the current catalog for availability and pricing.