Make π0.5 Learn to Call Astra for Reliable Robot Manipulation

Jianli Sun; Chengxiang Li; Yonglin Tian

cs.ROOct 9, 2026

Make π0.5 Learn to Call Astra for Reliable Robot Manipulation

Authors: Jianli Sun, Chengxiang Li, Yonglin Tian

Organizations: Institute of Automation, Chinese Academy of Sciences · School of Information and Intelligent Engineering, University of Sanya

Abstract

GPT-6 Astra provides robust embodied reasoning, but frequent external queries increase inference time and token costs. We equip a frozen $π_{0.5}$ policy with a lightweight CALL head that learns to predict when assistance is necessary. CALL reuses features from the standard policy forward pass and operates locally without consulting external models. Astra is activated solely for critical recovery events and returns control to $π_{0.5}$ upon stabilization. Evaluated across 40 LIBERO tasks (200 episodes), our system achieves a 97.5% success rate, outperforming the 95.5% baseline of pure $π_{0.5}$ . In the Object, Goal, and Spatial suites, Astra executes only 5.49% of environment steps, consuming approximately 5,027 tokens per test episode. This selective intervention preserves the fast ordinary execution of $π_{0.5}$ while seamlessly integrating Astra's reasoning on demand. Our project page is available at https://kykgeek.github.io/Pi05_CALL_Astra/.

Figures & tables

Branch	Input dimension	Projection dimension
Semantic representation	2,048	128
Action representation	1,024	128
Robot state and action statistics	44	64
Concatenated projection		320

Table 1: CALL-head branches. Each projection is a linear layer followed by GELU.

Suite	Held-out task folds	Competence-state examples
LIBERO-10	10	2,150
Object	10	1,625
Goal	8	860
Spatial	9	420
Total	37	5,055

Table 2: Offline competence-supervision data across task folds.

Figure 1: Method overview. (a) CALL reuses frozen $\pi_{0.5}$ features to predict when help is needed. Normal execution proceeds without an Astra query; selected requests activate Astra recovery, followed by a fresh $\pi_{0.5}$ proposal when control returns. (b) Counterfactual $\pi_{0.5}$ continuations provide soft help labels for CALL training. (c) Real wrist-camera frames show execution, a CALL trigger, Astra recovery, resumed $\pi_{0.5}$ control, and task completion within one episode.

Suite	Pure $\pi_{0.5}$	Astra + $\pi_{0.5}$	Difference
LIBERO-10	43/50 (86%)	47/50 (94%)	+8 pp
Object	50/50 (100%)	50/50 (100%)	0 pp
Goal	49/50 (98%)	49/50 (98%)	0 pp
Spatial	49/50 (98%)	49/50 (98%)	0 pp
Total	191/200 (95.5%)	195/200 (97.5%)	+2 pp

Table 3: Success over 50 episodes per suite and 200 overall. Differences are percentage points (pp).

Figure 2: Task success and actual Astra intervention across four LIBERO suites and overall. (a) Success rates of pure

\pi_{0.5}

and the hybrid system. (b) Percentage of test episodes containing executed Astra actions and percentage of environment steps executed by Astra.

Suite	Handoff ep.	Action ep.	Astra steps	Step share
LIBERO-10	31/50	9/50	3,485	20.43%
Object	19/50	1/50	132	1.86%
Goal	8/50	2/50	560	8.97%
Spatial	3/50	1/50	384	6.14%
Total	61/200	13/200	4,561	12.44%

Table 4: Astra handoff episodes, action episodes, and environment-step share.

Evaluation subset	Episodes or states	Success rate
Episodes without Astra consultation	139 episodes	139/139 (100%)
Episodes without executed Astra actions	187 episodes	187/187 (100%)
Pure $\pi_{0.5}$ at first CALL states	59 states	229/295 (77.63%)
Pure $\pi_{0.5}$ before actual Astra intervention	13 states	33/65 (50.77%)

Table 5: Success without assistance and pure

\pi_{0.5}

continuation success at assistance states. Each selected state has five continuation trials.

Suite	At first CALL states	Before actual Astra intervention
LIBERO-10	103/150 (68.67%)	25/45 (55.56%)
Object	91/95 (95.79%)	5/5 (100%)
Goal	25/40 (62.50%)	3/10 (30%)
Spatial	10/10 (100%)	0/5 (0%)
Total	229/295 (77.63%)	33/65 (50.77%)

Table 6: Pure

\pi_{0.5}

continuation success across suites. The CALL and pre-intervention columns use the selected states in Table 5 .

Suite	Input tokens	Output tokens	Total tokens	Tokens/test episode
LIBERO-10	2,278,625	143,566	2,422,191	48,443.82
Object	93,608	5,677	99,285	1,985.70
Goal	375,366	24,869	400,235	8,004.70
Spatial	238,278	16,248	254,526	5,090.52
Total	2,985,877	190,360	3,176,237	15,881.19

Table 7: Astra token expenditure across suites.

Figure 3: Token expenditure and measured action throughput. (a) Astra tokens per test episode on a logarithmic scale. (b) Environment actions per wall-clock second for pure

\pi_{0.5}

, hybrid episodes with CALL active and no Astra handoff, and all recorded hybrid episodes. Overall throughput is computed from summed steps divided by summed elapsed time.

Suite	Pure $\pi_{0.5}$	Hybrid no handoff	Hybrid all logged episodes
LIBERO-10	8.84	8.11	1.93
Object	8.54	8.35	4.06
Goal	8.17	8.25	2.97
Spatial	6.79	6.94	4.07
Overall	8.25	7.83	2.58

Table 8: Measured action throughput in environment steps per wall-clock second.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Suite	Task	Success	Handoff	Action ep.	Astra steps	Head
LIBERO-10	0	4/5	1	1	480	Fold
LIBERO-10	1	5/5	0	0	0	Fold
LIBERO-10	2	5/5	5	1	30	Fold
LIBERO-10	3	5/5	5	0	0	Fold
LIBERO-10	4	5/5	2	0	0	Fold
LIBERO-10	5	5/5	0	0	0	Fold

Appendix

Table A1: Task-level success and selective intervention.

Suite	Task	States	Success rate
LIBERO-10	0	1	5/5 (100%)
LIBERO-10	2	1	0/5 (0%)
LIBERO-10	7	2	10/10 (100%)
LIBERO-10	8	4	10/20 (50%)
LIBERO-10	9	1	0/5 (0%)
Object	6	1	5/5 (100%)

Appendix

Table A2: Task-level pure

\pi_{0.5}

continuation before actual Astra intervention.

Suite	Task	States	Success rate
LIBERO-10	2	5	25/25 (100%)
LIBERO-10	3	5	3/25 (12%)
LIBERO-10	4	2	10/10 (100%)
LIBERO-10	6	5	15/25 (60%)
LIBERO-10	7	4	20/20 (100%)
LIBERO-10	8	4	10/20 (50%)

Appendix

Table A3: Task-level pure

\pi_{0.5}

continuation at first CALL states.

Robot Policy Learning

Robot policy learning develops controllers that map sensory observations, task goals, and interaction history to actions for embodied robots.

See all 377 papers

An Unexpected Robot Policy: Early Evaluations of GPT-6 Astra on RoboDojo and Beyond
Sep 21, 2026Wenbo Zhang, Kaixuan Wang, Yutao Ouyang +9Language-Conditioned Robot Manipulation Generalization in Robotic Manipulation
RoboProbe · RoboDojo · The University of Hong Kong +5
Systematically Exploring the Capabilities of GPT-6 Astra as Embodied Policies
Sep 29, 2026Galbot Team, Xuchuan Chen, Xiaoqian Cheng +31Language Model-Based Control Robot Foundation Models
Galbot Team
Has GPT-6-Astra Achieved Embodied Intelligence?
Oct 15, 2026SenseNova, Joint Research Team, Zhongang Cai +59Embodied AI Agentic AI for Robotics

Sep 21, 2026J/K move · Enter open · S save

An Unexpected Robot Policy: Early Evaluations of GPT-6 Astra on RoboDojo and Beyond

Wenbo Zhang, Kaixuan Wang, Yutao Ouyang, Xiaoyu Huang+8

RoboProbe · RoboDojo · The University of Hong Kong+5

Language-Conditioned Robot Manipulation Generalization in Robotic Manipulation Robotic Policy Evaluation Robotic Manipulation Robot Manipulation Benchmarks

Open paper AskPDF arXiv

Embodied AI systems are often organized into System 1 and System 2. System 1 is typically a pretrained policy that generates actions at high frequency, whereas System 2 is often instantiated as a vision-enabled language model for high-level planning. We ask whether a large language model (LLM) can act as the policy for robot manipulation without task-specific finetuning. We call this setting LLM as policy. We evaluate three LLMs on all 42 RoboDojo tasks and compare their scores with 40 public policies. Astra and GPT-5.5 use the official 50-episode-per-task protocol; DeepSeek-Flash uses 10 episodes per task. GPT-6 Astra achieves 22.48% average success rate and 28.97 Score over 2,100 trials, ranking above every public entry. Yet GPT-5.5 and DeepSeek-Flash reach only 0.88% and 1.92% average success rate with the same post-processing. We find that Astra exhibits a sharply polarized capability profile. It generalizes well to tasks that require semantic understanding but not high-precision control. In contrast, it performs poorly on tasks that require precision, dynamic control, or complex bimanual coordination. In-context experiments show no aggregate benefit from one-shot demonstrations, while selected interaction traces show within-episode corrections under perturbations. Overall, the evaluated LLMs vary substantially in manipulation performance. Astra stands out and provides initial evidence for the potential of a general-purpose manipulation model, although reliable precision and dynamic control remain limitations in the evaluated setting.

Abstract

Robot Policy Learning

Explore similar work

An Unexpected Robot Policy: Early Evaluations of GPT-6 Astra on RoboDojo and Beyond

Systematically Exploring the Capabilities of GPT-6 Astra as Embodied Policies

Has GPT-6-Astra Achieved Embodied Intelligence?