ZTE's EmbodiedMind Cracks Long-Horizon Robot Planning with Trie-GRPO

Long-Horizon Planning: The New Watershed for Embodied Intelligence

2026 marks a turning point where embodied intelligence moves from demos to real-world deployment. According to the China Electronics Society, domestic humanoid robot shipments exceeded 40,000 units in H1 2026, accounting for 97% globally. OpenAI's CEO confirmed plans to develop humanoid robots, emphasizing that the robot's "brain" matters more than its form, targeting a 2027 prototype. The industry consensus: morphology is no longer the differentiator; the embodied brain's planning, reasoning, and generalization in complex scenes determine whether a robot can actually work.

Long-horizon task planning — moving from single-step recognition to multi-step continuous decision-making, from "understanding the scene" to "decomposing tasks, planning paths, and correcting errors" — is the key leap for embodied brains. ZTE's paper EmbodiedMind: Adaptive Data Curation and Prefix-Tree Reinforcement Learning for Efficient Embodied Intelligence (submitted to ICRA 2027, arXiv:2609.19659) addresses this with an 8B model scoring 70.02% average on 18 embodied-brain benchmarks, improving long-horizon planning accuracy by 30.31% over prior best methods.

Three Structural Dilemmas in Embodied Brain Training

Current embodied foundation model training suffers from three structural defects:

Low sample efficiency: Training corpora contain many low-information-density, noisy trajectories. Supervised fine-tuning and RL waste compute on redundant, high-variance data, converging to suboptimal policies while valuable signals are diluted.

Heterogeneous task gradient imbalance: Dynamic sampling methods like DAPO lack task-level calibration. Simple tasks are over-discarded, hard tasks systematically ignored, causing a "seesaw effect" where capabilities trade off against each other in multi-task training.

Distorted credit assignment in long-horizon planning: Standard GRPO applies trajectory-level advantage uniformly to all tokens. A failed trajectory penalizes even its correct intermediate decisions. Tree search can backtrack to single-step nodes but faces combinatorial explosion, struggling to balance exploration depth and computational feasibility.

These three dilemmas point to a core proposition: the training paradigm itself must be reconstructed for embodied brains to move from "understanding environments" to "long-horizon execution."

Three Mechanisms for Systematic Breakthrough

The team proposes a three-stage training paradigm: RSFT for data curation and SFT, IR-GRPO for task balance, and Trie-GRPO for step-level credit assignment. Each mechanism targets one dilemma, progressing layer by layer.

Mechanism 1: RSFT — Boosting Data Information Density at Source

To address low-information-density trajectories diluting effective signals, RSFT uses high-temperature repeated inference to stratify samples: all-correct → simple, partially correct → uncertain, all-failed → hard. The training set retains only uncertain samples and a proportion of simple samples, temporarily excluding hard samples. This prevents early distribution collapse and avoids contamination from erroneous trajectories, raising data information density at the source. Empirically, RSFT outperforms standard SFT across multi-dimensional capability evaluations.

Mechanism 2: IR-GRPO — Rebalancing Heterogeneous Tasks

To fix GRPO's lack of task-level calibration causing capability trade-offs, IR-GRPO maintains an independent FIFO queue per task type and uses exponential moving average to smooth each task's valid sample ratio, dynamically adjusting the next round's sampling quota. This keeps data proportions balanced across tasks during training. Combined with a dedicated reward mechanism for diverse embodied tasks, experiments show IR-GRPO improves overall model capability while significantly reducing training time.

Mechanism 3: Trie-GRPO — Step-Level Credit Assignment

The core insight: the policy optimization "group" should not be defined at the full trajectory level but conditioned on action prefixes. All sampled trajectories are inserted in parallel into an action prefix tree (trie). Trajectories sharing the same history prefix form independent decision subgroups at branching points, and relative advantage is computed strictly among sibling nodes.

This means even if a trajectory ultimately fails, its intermediate correct actions still receive positive local advantage, eliminating the "collective punishment" problem. Correct vs. incorrect actions at branching points are directly compared, yielding step-precise advantage signals. To control exploration cost, discounted returns are back-propagated along tree paths, sibling-node advantages are normalized, and near-zero advantage samples are filtered, achieving a principled balance between exploration depth and computational feasibility.

From 18 Benchmarks to Physical World: Data-Driven Proof

EmbodiedMind-8B uses only 522,000 training samples to achieve 70.02% average on 18 embodied-brain benchmarks, surpassing same-scale open-source models. The 18 benchmarks cover four capability blocks: general multimodal understanding, 2D spatial reasoning, 3D spatial measurement & task planning, involving counting, object localization, distance/size measurement, planning step generation, action prediction, etc., comprehensively testing the spectrum from environment understanding to task execution.

In simulation, the model scores 66.67 on VLM-PlanSim-99 (max 25-step planning), beating Cosmos3-Nano to reach industry-leading level. On ALFWorld-thor it achieves 60.45% success rate , exceeding SEEA-R1's 46.27%.

Real-robot validation is even more compelling. Using the AgiBot A2 robot, the team built 50 desktop scenarios and collected 160 task execution monitoring videos as the A2-Failure dataset. EmbodiedMind reaches 62.00% on desktop tidying planning , far above Qwen3-VL-8B's 40.00% baseline; task execution monitoring accuracy is 55.31% , also leading Qwen3-VL-8B's 46.88%. This demonstrates that step-level credit assignment learned via Trie-GRPO in simulation transfers effectively to real robots — long-horizon planning is no longer just a benchmark score but yields observable success-rate gains in the physical world.

Leaderboard Validation: Entering the Domestic First Tier

Third-party evaluation provides external verification. SuperCLUE's September EmbodiedCLUE-VLA benchmark covers 10 representative models across basic perception, visual reasoning, interaction & planning, and embodied safety. ZTE's Nebula-EmbodiedBrain-180B scores 77.48 , tying with the 2.4T-parameter Qwen3.8-Max-0902 for domestic first place, entering the global first tier.

On the embodied safety sub-leaderboard, Nebula-EmbodiedBrain leads with 97.92 , showing high reliability and safety. With 180B parameters, it aligns performance with trillion-parameter super-models. International reference models GPT-6 Astra (86.75) and Gemini-3.8-Flash (82.12) are not ranked domestically; ZTE's 77.48 places it domestic co-first and global first tier.

Complete Loop: From Bottom-Up Paradigm to Top-Level Capability

From solving the "collective punishment" problem in long-horizon planning, to outperforming same-scale open models with minimal samples, to tying with international giants on authoritative third-party leaderboards, this training paradigm completes a full loop from "bottom-layer training paradigm" to "top-layer general capability" in embodied intelligence.

Solid methodology first, hard leaderboard results second — this not only validates the technical route's advancement and generalization but also provides the industry with a reproducible, cost-effective embodied brain R&D path. As embodied intelligence accelerates into the physical world, robot brains capable of "finishing a task in one breath" are injecting strong momentum for robots to truly enter thousands of industries.

References: [1] EmbodiedMind: Adaptive Data Curation and Prefix-Tree Reinforcement Learning for Efficient Embodied Intelligence, arXiv:2609.19659 [2] SuperCLUE 2026 September EmbodiedCLUE-VLA (Embodied Brain) Evaluation Leaderboard