Has GPT-6-Astra Achieved Embodied Intelligence?

SenseNova, Joint Research Team, :, Zhongang Cai, Ziang Cao, Tailai Chen, Han Chong Chua, Linjun Dai, Hanming Deng, Hokin Deng, Haiwen Diao, Chenjing Ding, Weichen Fan, Xiao Fu, Zerui Gong, Chenyang Gu, Zexin He, Zhi Hou, Bo Li, Qinyue Li, Quanzhou Li, Xiaowan Li, Zhichao Liao, Dahua Lin, Zilong Lin, Bo Liu, Chunxiao Liu, Pengfei Liu, Tianqi Liu, Ziwei Liu, Chen Change Loy, Lewei Lu, Jiapeng Luo, Yuexiao Ma, Liang Pan, Xingang Pan, Oscar Qian, Ziyi Shang, Yuyao Shao, Dawei Su, Dacheng Tao, Qingyi Tao, Quan Wang, Ruisi Wang, Xiaogang Wang, Ziyi Wang, Beichen Wen, Donghao Wu, Jiaxu Wu, Silei Wu, Xiaoming Wu, Zhonghua Wu, Hongcan Xiao, Weiye Xu, Atsushi Yamashita, Lei Yang, Zhe Yang, Baiqiao Yin, Wanqi Yin, Hanfei Zhao, Tianbo Zhong, Long Zhuo, Jing Zuo · N/A · 2026

Summary

GPT-6-Astra demonstrates a level of competence in robotic manipulation and navigation that challenges the conventional paradigm of building specialized models for individual tasks and embodiments.

Abstract Summary

GPT-6-Astra demonstrates a level of competence in robotic manipulation and navigation that challenges the conventional paradigm of building specialized models for individual tasks and embodiments. In this report, we ask not only how capable GPT-6-Astra is, but what its capabilities imply for the field. First, we probe the limits of standalone GPT-6-Astra. We conduct a large-scale evaluation across 12 representative benchmarks spanning tabletop manipulation, humanoids, mobile platforms, and dexterous hands. Under a unified protocol that allows in-context learning (ICL) from training-set demonstrations and substantially increases the number of execution steps, GPT-6-Astra achieves unprecedented performance (Fig. 1). Future research may focus on improving 1) the generality of ICL 2) cost and execution time, 3) physical-state verification, 4) recovery from irreversible failures, and 5) long-horizon preservation of progress. Second, we study the agentic harness for embodied control. We identify that close-loop coordination between a high-level "System 2" reasoner (i.e., GPT-6-Astra) and a low-level "System 1" controller (e.g., {\pi}0.5) is costly but effective. Third, we examine GPT-6-Astra on real-world deployment. Our experiments on 8 robot configurations reveal substantial gaps: 1) robot demonstrations required for ICL are expensive to collect, and 2) it is implausible to trade longer execution for higher success rates with physical constraints. Fourth, we envision the next generation of embodied benchmarks. We develop a prototype that 1) derives tasks from real-world needs, 2) provides no in-domain demonstrations, and 3) judges agents on a world clock that keeps running while they deliberate. Taken together, our results suggest that GPT-6-Astra has largely crossed the threshold of embodied competence, but not yet that of practical embodied intelligence.

Key Points

GPT-6-Astra demonstrates a level of competence in robotic manipulation and navigation that challenges the conventional paradigm of building specialized models for individual tasks and embodiments.
In this report, we ask not only how capable GPT-6-Astra is, but what its capabilities imply for the field.
First, we probe the limits of standalone GPT-6-Astra
We conduct a large-scale evaluation across 12 representative benchmarks spanning tabletop manipulation, humanoids, mobile platforms, and dexterous hands.
Under a unified protocol that allows in-context learning (ICL) from training-set demonstrations and substantially increases the number of execution steps, GPT-6-Astra achieves unprecedented performance (Fig.

Large-Scale Study under A Unified Protocol

We study GPT-6-Astra across twelve embodied settings, spanning manipulation, humanoid loco-manipulation, household activities, and navigation, under a single protocol in which its model weights remain fixed.

Two conditions shape the study: in-context learning (ICL), which gives GPT-6-Astra access to the same training demonstrations from which conventional policies are learned, and an extended interaction budget, which gives it more opportunity to correct mistakes within an episode.

Large-Scale Study under A Unified Protocol — Has GPT-6-Astra Achieved Embodied Intelligence? — Figure 1 : Probing the capability upper bound of GPT-6-Astra across 12 benchmarks under a unified protocol that 1) allows for in-context learning (ICL) on the training set and 2) extends the step limit (budget) up to $10\times$. RoboLab has no training set and thus no ICL; BEHAVIOR is not extended to $10\times$ due to its large official budget. SOTA cut-off date: October 12, 2026. Details in Tab. 2 .

Harness Ablation

Section 2.2 shows that additional interaction can substantially improve task completion, but its benefits depend on how effectively the agent uses that interaction. Longer execution incurs substantial inference costs and does not reliably resolve failures in physical verification, recovery, or preservation of subtask progress.

These findings motivate examining the harness that connects GPT-6-Astra’s reasoning to observations, actions, and feedback.

Harness Ablation — Has GPT-6-Astra Achieved Embodied Intelligence? — Figure 2 : In-context learning protocol. For each task, GPT-6-Astra studies the training data and writes a knowledge package $K$, which is frozen before testing. Each test episode starts a fresh session with $K$ and runs under the standard ($1\times$) or extended ($10\times$) budget. Model weights remain fixed throughout, and $K$ remains unchanged during testing, and no information passes between episodes.

Real-World Validation

Since the simulation protocol of Sec. 2.1 rests on two assumptions that do not hold on physical robots, namely that successful demonstrations of the test task are available and that additional execution steps are free, we further study a real-world variant of the protocol that revises both.

First, the context changes: since new tasks often come without demonstrations in real-world deployment, GPT-6-Astra operates in a zero-shot setting and receives only structured information provided by a human as context.

Real-World Validation — Has GPT-6-Astra Achieved Embodied Intelligence? — Figure 3 : Contents of an ICL knowledge package. GPT-6-Astra decides what to retain, yet packages share six recurring components; file types are typical and content varies by benchmark. At test time, $K$ is read alongside current observations to retrieve, adapt and check actions.

Towards a Next-Generation Embodied Benchmark

The preceding sections show that GPT-6-Astra’s results on established embodied benchmarks depend heavily on conditions that the benchmarks grant. Its success rises with in-context access to the benchmark’s demonstrations and with an extended step budget in which failures cost nothing (Sec. 2 ), and depends on how a general agent is adapted to an action interface built for a learned policy (Sec. 3 ).

Contributors

Zhongang Cai*, Zhonghua Wu*, Qingyi Tao*. Simulation Benchmarks: Quanzhou Li*, Han Chong Chua*, Ruisi Wang*, Xiaoming Wu*, Tianqi Liu*, Bo Liu*, Pengfei Liu*, Jing Zuo*, Hongcan Xiao*, Qinyue Li. Harness Framework: Weiye Xu*, Yuyao Shao*, Silei Wu*, Jiapeng Luo*, Hanming Deng*, Tailai Chen*, Linjun Dai*, Ziyi Shang*, Xiaowan Li, Donghao Wu, Tianbo Zhong, Chenjing Ding.

Full Simulation Results

Tasks and demonstrations. LIBERO-PRO [ 183 ] applies task-preserving perturbations to the four LIBERO suites and separates each suite into Task and Swap variants. We evaluate the complete official benchmark: eight suites, ten tasks per suite, and ten fixed initial states per task, for 800 episodes in total.

Before evaluation, we freeze a package selected by the exact base task, combining matching Task/Swap task-family notes with global strategy, perception, verification, and recovery notes distilled from a separate public exploration phase.

Full Simulation Results — Has GPT-6-Astra Achieved Embodied Intelligence? — Figure 6 : Residual failures under ICL with the extended budget. Each pair shows the state before a key event and its outcome; purple arrows mark the relevant objects. (a) Verifying physical success: apparent or self-reported completion diverges from the actual task state. (b) Recovering from consequential actions: an early mishap, such as a spill or a dropped part, cannot be undone and derails later subtasks. (c) Preserving progress over long horizons: progress that later steps depend on, such as an open lid or a placed object, is lost and blocks the task.

Harness Ablation Study

Configuration Average Generalization Precision Long-Horizon Memory Open Tokens (M) $\downarrow$ Baseline 22.78 / 15.56 27.08 / 22.22 9.79 / 0.00 23.13 / 4.17 22.22 / 22.22 31.67 / 29.17 2.56 A.

Agent Components w/ planning subagent 26.03 / 19.44 37.08 / 33.33 11.88 / 4.17 21.46 / 4.17 40.56 / 38.89 19.17 / 16.67 2.96 w/ execution subagent 21.56 / 15.28 36.39 / 33.33 14.79 / 4.17 10.63 / 0.00 25.56 / 22.22 20.42 / 16.67 2.55 w/ feedback subagent 30.00 / 24.72 30.83 / 27.78 16.88 / 12.50 24.38 / 8.33 33.33 / 33.33 44.58 / 41.67 2.74 w/ history co...

Harness Ablation Study — Has GPT-6-Astra Achieved Embodied Intelligence? — Figure 7 : Token cost versus performance on RoboDojo for different System 1–System 2 collaboration strategies and action interfaces. Each point represents an experimental configuration, with token cost measured in millions and performance measured by the unweighted mean of category scores. Dashed lines intersect at the Agentic GPT-6-Astra baseline reference point, defining four shaded quadrants.

Full Real Robot Results

This appendix documents the real-world experiments of Sec. 4 . App. 11.1 summarizes all platforms, and App. 11.2 – 11.9 document each platform, from the simplest to the most complex embodiment, in the same five parts: setting, context provided, tasks and results, findings, and visualization. Table 14 summarizes the cameras, action interfaces, supplied context, and task outcomes across all eight platforms.

Figure 13 provides task visualizations grouped by outcome, while the following subsections detail each platform’s protocols and results.

Full Real Robot Results — Has GPT-6-Astra Achieved Embodied Intelligence? — Table 1 : Real-world platforms, ordered from simple to complex embodiments. Each class probes a different question; settings and per-task results are in App. 11 .

Sources and Demonstrations

Figures are reproduced from SenseNova et al. See the full paper for experimental details and the project page, when the paper links one, for demonstrations and videos.