Literature
Display
← All papers

G0.5: One Autoregressive Stream for Robot Reasoning and Action

Yicheng Liu · Zibin Dong · Baijun Ye · Tianyuan Yuan · Tao Jiang · Anqi Yang · Shicheng Cao · Haonan Liu · Yue Sun · Zihan Guo · Xiao Liu · Dong Ke · Changxun Pan · Chenru Wu · Tailai Cheng · Xiaoshu Ren · Xinlei Zhang · Jianning Cui · Zijie Zhao · Haoyu Zhang · Kaiming Xu · Haodong Yang · Bowen Zhang · Jiahui Niu · Shaoting Zhu · Shiduo Zhang · Hang Zhao

Preprint posted · arXiv (preprint) ↗

Robotics14 min text · CondensedOriginal paper ↗
Contents

Overview

Complete original PDF block S0.F1
Figure 1: G0.5: reasoning and action in one autoregressive stream. Top left: A single VLM conditions on multi-view images, an instruction, proprioception, and an embodiment identifier, then generates optional chain-of-thought (CoT) and compact action codes under the same next-token objective. A cross-embodiment ActionCodec converts the codes into continuous motor commands, while active-part tokenization omits inactive control groups. Top right: Comparisons with flow-matching policies span generalist pretraining and specialist post-training, covering optimization, compatibility with reinforcement learning, transfer after DROID post-training to a held-out physical setup and previously unseen object instances, and inference latency. Bottom: A closed-loop rollout of open the drawer →\rightarrow pour the walnuts →\rightarrow close the drawer. At each stage, updated visual and proprioceptive observations condition interleaved reasoning and action generation. The examples highlight prompt-driven action steering (left), grounded bounding-box prediction (middle), and adaptive subtask reasoning (right); colors distinguish conditioning context, reasoning, and action tokens.

At a glance

This edition follows all 32 pages of Galaxea G0.5 v 1, preserving twelve figures, seven numbered tables, the complete token template, action layout, displayed objective and references. A Qwen3.5-2B decoder emits optional reasoning and discrete action codes; a separately learned ActionCodec converts codes into continuous robot commands.

Reported results include 76.7% real-task success across six settings, 82.5% held-out DROID score after DROID post-training, 98.9% LIBERO success and 0.3136 BEHAVIOR partial-progress score after four epochs. These are distinct protocols and metrics, not interchangeable measures of general capability.

CoT benefits the tested stage-conditioned long tasks more than single-stage picking. Pretraining data, embodiment exposure, auxiliary codec and optional flow head remain parts of the system; the experiments do not isolate an architectural cause for every gain.

1 Introduction

Instead of using a vision-language model only to condition a continuous action expert, G0.5 predicts reasoning and actions in a shared autoregressive vocabulary. Grouped action compression reduces token count, optional CoT supplies intermediate grounding, and short visual history enters the encoder.

Prompt edits can influence actions in either conditioned or shared-decoder systems; the proposed distinction is the direct token-generation interface, not a general impossibility of influencing an external expert. Figure 1 retains the complete rollout and latency panel: FlashRT AR 130 ms without CoT or 192 ms with CoT versus FM 190 ms. Hardware and a detailed timing protocol are not specified in this report.

Complete original PDF block S1.F2
Figure 2: Token sequence template. All inputs and outputs are serialised into a single autoregressive sequence: a conditioning segment (multi-view RGB, embodiment id, task instruction, proprioceptive state—in user-side chat tokens) and a generative segment on which the next-token cross-entropy loss in Eq. (1) is applied. The generative segment composes an optional chain-of-thought span—any subset of four self-describing reasoning targets (Subtask:, BBox:, Trace:, ActionHint:)—followed by the action codes, which themselves expand into RR residual rounds of DoF-group markers each followed by 8 action codes (Sec. 3.1).

2 Related Work

2.1 VLA Architectures: from VLM-as-Encoder to VLM-as-Actor

The paper contrasts autoregressive action-token policies such as RT-2 and OpenVLA with continuous experts such as pi0. Separate action losses can change pretrained features; shared token prediction preserves the original objective form, but does not by itself prove that language capabilities cannot degrade.

2.2 Action Tokenization and Cross-Embodiment

FAST compresses trajectories through a fixed transform, while ActionCodec and FASTer motivate learned residual codes and grouping. G0.5 maps supported bodies into five fixed parts and 27 continuous dimensions, then emits sparse group tokens. Adding a body without new head parameters presumes compatibility with that representation, not arbitrary morphology.

2.3 Reasoning and Chain-of-Thought in VLAs

HAMSTER separates high-level reasoning from control; ECoT trains reasoning before actions. G0.5 combines subtask text, object boxes, 2 D gripper traces and action hints in prompt-selectable formats. The relevant novelty claim concerns this combination with grouped actions and a shared decoder.

3 G0.5 Model Design

Inputs are multi-camera RGB history, embodiment identifier, task and proprioception. Qwen3.5-2B predicts the assistant segment, and Equation 1 applies next-token cross-entropy only to those generated tokens. Conditioning tokens are masked from the loss.

The default policy is autoregressive; an optional pi0.5-style flow-matching head conditions on trunk states for comparisons. The common decoder loss does not include the separate action codec’s earlier training objectives. Figure 2 preserves the full sequence format and all special-token definitions.

Equation 1
ℒ(θ)=−∑i∈𝒢logpθ(xi∣x<i),\mathcal{L}(\theta)\;=\;-\!\!\sum_{i\in\mathcal{G}}\log p_{\theta}\bigl(x_{i}\mid x_{<i}\bigr), (1)

3.1 Structured Tokenization of Heterogeneous Action Data

Robot actions are divided into semantic motion parts, padded to a shared maximum dimension and encoded by a residual vector quantizer. A temporal contrastive term encourages adjacent motions to have consistent codes. Each active group contributes a marker and eight codes per residual round; inactive groups need no action tokens.

Figure 2 describes sparse active groups but its legend calls arm markers always emitted; Figure 3 also depicts separate gripper markers. Exact illustrations are retained without inventing a resolved serialization convention or an unspecified fixed residual-round count.

Complete original PDF block S3.F3
Figure 3: Structured action tokenization. Heterogeneous robot actions are decomposed into semantically aligned motion parts, encoded with a residual vector quantizer, and serialized as part-specific action tokens. This representation shares one action vocabulary across embodiments while allowing sparse prediction over only the activated parts.

3.2 Native Chain-of-Thought

Subtask, BBox, Trace and ActionHint fields can precede actions. Training samples eight curated formats, including no CoT, rather than all sixteen mathematical subsets of four optional fields. The source’s general “any subset” wording should be read with this restricted training recipe.

CoT is generated intermediate supervision, not verified internal reasoning. Its effect on behavior is evaluated by turning formats on/off at inference; the long-task probes still receive external stage instructions.

3.3 Visual Memory

Following pi0.7 and MEM, factorized temporal/spatial attention is inserted every four vision-transformer layers. Final-layer history tokens are discarded, and all historical frames are dropped on 30% of training examples. Continuous state embeddings align proprioception with frames.

This bounds the decoder’s visual-token input while retaining only a multi-second history. Single-frame BEHAVIOR evaluation cannot demonstrate online tracking of previous decisions; the proposed benefit there is a learned visual prior, not persistent episode memory.

4 G0.5 Pre-training

A single main pretraining stage mixes robot demonstrations and web/VQA samples under generated-token cross-entropy. Separate action-codec training precedes this stage. The report describes composition and sampling qualitatively but omits a full source-hour inventory, exact mixture weights and a reproducible total pretraining budget.

Robot data mixture.

Fourteen real/simulated embodiments share the exact 27-dimensional layout retained below:9 left control, 1 left gripper, 9 right control, 1 right gripper and 7 lower-body dimensions. Unused slots are padded at data merge; sparse token emission is a separate representation step. Normalization/clipping and filters vary by source.

DROID is excluded from foundation pretraining but used subsequently for its evaluation. Figure 4 shows long-tailed concepts dominated by picking/placing and household objects; occurrence frequencies are not a measured task-competence distribution.

Complete original PDF block source-action-layout
Exact 27-dimensional action layout
Original selectable source block
<left_control>(9) | <left_gripper>(1) | <right_control>(9) | <right_gripper>(1) | <lower_body>(7)
Complete original PDF block S4.F4
Figure 4: Action and object concept distribution in the pre-training corpus. We show the most frequent action verbs and object nouns extracted from the pre-training data; each bar gives the concept’s share of occurrences (%) on a logarithmic axis, and the trailing Others bar aggregates the remaining vocabulary.

Autolabeling pipeline.

Rule-based segmentation and keyframes feed multimodal APIs including Gemini3 and Doubao Seed2.0Pro for episode/subtask/action-hint labels. Foundation-model grounding plus SAM3 tracking supplies boxes/masks. Forward kinematics and camera projection supply 2 D end-effector traces; label accuracy is not independently quantified.

Web and VQA co-training.

Generic web VQA, embodied VQA and in-house subtask/grounding/commonsense labels are mixed with actions in an action-heavy ratio. Shared cross-entropy allows all targets in one vocabulary, although exposure counts and token-length differences can change their effective optimization weights.

Chain-of-thought supervision.

Each robot example receives one of eight weighted formats; subtask text is emphasized. The listed families include task text, subtask/action hints, trajectories and bounding boxes. Standard evaluations use no CoT, except the explicit CoT probes in Section 5.6.

Implementation Details.

AdamW uses warmup, a constant phase and cosine decay with the vision tower unfrozen. Sparse multi-second image history includes the present frame and stochastic history dropout. Exact base learning rate, steps, batch size, history offsets and total compute are not supplied here; downstream settings are more specific.

5 Experiments

Evaluation separates held-out environment/object transfer after dataset adaptation, simulation fine-tuning, long-horizon mobile progress, physical fine-tuning, referring-context effects and inference-time CoT/head switching. Claims of zero-shot performance must be interpreted relative to each stated training exclusion.

5.1 DROID Environment- and Object-Level Zero-Shot Evaluation

G0.5 is post-trained on DROID before deployment. Only the evaluation room and physical object instances are excluded from both training stages. This is environment/object-level transfer, not zero-shot transfer to the DROID dataset.

5.1.1 Evaluation Setup

Robot Platform.

The platform is a 7 DoF Franka Research 3 with a Robotiq 2 F-85 gripper, desk mount, right-side global RGB camera and wrist RGB camera. Instructions and both views condition control.

Tasks.

Ten tasks cover carrot/peach to bowl, block to green/red plate or cup, towel/pen to drawer, moving a bowl left, towel bowl-to-plate and block insertion followed by drawer closure. Figure 5 retains all eight scene panels and their labels.

Evaluation Protocol.

Each task has ten trials. Nine use binary completion; block-in-drawer-plus-close awards 0.5 for insertion only. Thus the reported 82.5% average includes partial credit and is not simply the fraction of 100 fully successful trials.

Baselines.

Comparators are pi0.5-DROID and MolmoAct 2-DROID, with different foundation models and preprocessing. Their common downstream dataset does not equalize pretraining data or isolate decoder architecture.

Complete original PDF block S5.F5.sf1
(a) Carrot / Peach →\rightarrow Bowl
Complete original PDF block S5.F5.sf2
(b) Block →\rightarrow Green / Red Plate
Complete original PDF block S5.F5.sf3
(c) Block →\rightarrow Cup
Complete original PDF block S5.F5.sf4
(d) Towel →\rightarrow Open Drawer
Complete original PDF block S5.F5.sf5
(e) Pen →\rightarrow Open Drawer
Complete original PDF block S5.F5.sf6
(f) Bowl →\rightarrow Left
Complete original PDF block S5.F5.sf7
(g) Towel: Bowl →\rightarrow Plate
Complete original PDF block S5.F5.sf8
(h) Block →\rightarrow Drawer + Close
Figure 5: DROID environment- and object-level zero-shot evaluation tasks. After DROID post-training, we evaluate G0.5 on 10 manipulation tasks across 8 held-out scene setups and previously unseen physical object instances on a Franka Research 3 robot arm. Tasks cover object placement, color-conditioned target selection, small-aperture insertion, deformable object manipulation, spatial displacement, and multi-step sequential execution.

5.1.2 Quantitative Results

The reported aggregate scores are 82.5% G0.5, 57.5% pi0.5 and 52.0% MolmoAct 2:25.0 and 30.5 percentage-point gaps. Figure 6 preserves every task bar. Small per-task samples and the lack of intervals limit precision.

G0.5 -DROID vs. π 0.5 -DROID.

The main marked-scene comparison reports G0.5 ahead of pi0.5 on all ten tasks. This statement does not extend to the earlier unmarked drawer test, where G0.5 scored lower.

G0.5 -DROID vs. MolmoAct2-DROID.

MolmoAct 2 is reported to freeze near some objects or grasp before reaching them, and scores zero on the combined insertion/closure task. These qualitative failure observations complement the tested scores without establishing a general model limitation.

Effect of visual contrast on drawer localisation.

Without orange drawer markers, towel insertion scores were 90% pi0.5, 80% MolmoAct 2 and 60% G0.5. Adding high-contrast cards raises G0.5 to 100%. The marker-assisted main result therefore depends on a visual intervention that particularly helps G0.5; semi-transparent target localization remains a weakness.

Complete original PDF block S5.F6
Figure 6: DROID environment- and object-level zero-shot evaluation results. All models are trained on DROID data, while the physical evaluation environment and object instances are held out. Per-task success rates (%) are reported for π0.5\pi_{0.5}-DROID, MolmoAct2-DROID, and G0.5 across 10 manipulation tasks. G0.5 achieves an average of 82.5%, outperforming π0.5\pi_{0.5}-DROID (57.5%) by 25.0 percentage points and MolmoAct2-DROID (52.0%) by 30.5 percentage points.

5.2 Simulation Benchmarks

5.2.1 Bridge-SimplerEnv

BridgeDataV2 fine-tuning uses 80 ksteps at 3×10⁻⁵ and no proprioception in either training or evaluation. Four SimplerEnv WidowX tasks score 97.5, 75.0, 83.3 and 93.3%, averaging 87.3%, with no extra simulation-domain training. Baselines are compiled from prior benchmark studies, not all rerun under an identical training budget.

Complete original PDF block S5.T1
Table 1: Results on Bridge-SimplerEnv. All numbers are success rates (%). Since several original model papers do not report results on SimplerEnv-WidowX, we compile the corresponding baseline numbers from prior studies that evaluate these models on this benchmark. The best and second-best results are highlighted in bold and underlined, respectively.
Original selectable table
Method
Put Spoon
on Towel
Put Carrot
on Plate
Stack Green Block
on Yellow Block
Put Eggplant
in Yellow Basket
Average
π0\pi_{0} [3] 29.1 0.0 16.6 62.5 27.1
π0\pi_{0}-FAST [11] 29.1 21.9 10.8 66.6 32.1
π0.5\pi_{0.5} [4] 49.3 64.7 44.7 69.7 57.1
GR00T-N1.5 [5] 75.3 54.3 57.0 61.3 61.9
StarVLA-GR00T [46] 83.0 59.4 18.8 100.0 65.3
RoboBrain2.5-8B [42] 75.0 55.5 40.1 100.0 67.6
MemoryVLA [47] 75.0 75.0 37.5 100.0 71.9
EO-1 [48] 63.6 54.5 81.8 90.9 72.7
Xiaomi-Robotics-0 [49] 95.8 62.5 75.0 83.3 79.2
G0.5 (Ours) 97.5 75.0 83.3 93.3 87.3

5.2.2 RoboTwin 2.0

RoboTwin fine-tuning combines 2, 500 clean and 25, 000 randomized demonstrations for four epochs, learning rate 4×10⁻⁵, batch 1, 024. The protocol states 100 trials/task/setting. Clean 93.72% and randomized 92.84% average 93.28%, rounded 93.3%; all fifty task rows remain in Table 7.

Complete original PDF block S5.T2
Table 2: Results on RoboTwin 2.0. We report success rates under clean and randomized evaluation settings, together with their average. The best and second-best results are highlighted in bold and underlined, respectively.
Original selectable table
Method Clean Rand. Average
π0\pi_{0} [3] 65.9 58.4 62.2
π0.5\pi_{0.5} [4] 82.7 76.8 79.8
LingBot-VLA [52] 86.5 85.3 85.9
Motus [50] 88.7 87.0 87.8
StarVLA [46] 88.7 87.8 88.3
LingBot-VA [51] 92.9 91.5 92.2
Fast-WAM [16] 91.9 91.8 91.8
G0.5 (Ours) 93.7 92.8 93.3

5.2.3 LIBERO

LIBERO has four ten-task suites with fifty demonstrations per task. Fine-tuning uses 100 ksteps, learning rate 10⁻⁵ and weight decay 10⁻², followed by fifty rollouts/task. Spatial 98.4, Object 100, Goal 98.6 andLong 98.6 average 98.9%. Differences near this ceiling are small and lack uncertainty estimates.

Complete original PDF block S5.T3
Table 3: Results on LIBERO. The best and second-best results are highlighted in bold and underlined, respectively.
Original selectable table
Method Spatial Object Goal Long Average
Qwen-VLA-Base [53] – – – – 90.8
π0\pi_{0} [3] 98.0 96.8 94.4 88.4 94.4
InternVLA-M1 [54] 98.0 99.0 93.8 92.6 95.9
Wall-OSS-0.5 [55] – – – – 96.5
π0.5\pi_{0.5} [4] 98.8 98.2 98.0 92.4 96.9
GR00T-N1.7 [5] 97.7 98.5 97.5 94.4 97.0
OpenVLA-OFT [56] 97.6 98.4 97.9 94.5 97.1
Fast-WAM [16] 98.2 100.0 97.0 95.2 97.6
Being-H0.5 [23] 99.2 98.2 99.0 96.2 98.2
Motus [50] 96.8 99.8 96.6 97.6 97.7
Qwen-VLA-Instruct [53] – – – – 97.9
EO1 [48] 99.7 99.8 99.2 94.8 98.4
Cosmos Policy [57] 98.1 100.0 98.2 97.6 98.5
LingBot-VA [51] 98.5 99.6 97.2 98.5 98.5
Xiaomi-Robotics-0 [49] 98.8 100.0 98.8 97.2 98.7
G0.5 (Ours) 98.4 100.0 98.6 98.6 98.9

5.3 Long-horizon Tasks

The 2025 BEHAVIOR challenge provides 10, 000 demonstrations across fifty household tasks, over 1, 100 hours, with ten evaluation instances per task. Its score is the fraction of satisfied BDDL goal predicates in the best matching goal clause, so 0.3136 means partial progress rather than 31.36% whole-episode completion.

Episodes average 6.6 minutes and can last 14 minutes, requiring R1-Pro navigation and bimanual manipulation. This is a simulator benchmark, distinct from the real six-setting evaluation.

5.3.1 Implementation Details

All tasks share one checkpoint, low-resolution RGB and single-frame post-training/evaluation, without the competition winner’s stage head. G0.5 and pi0.5 scores average two evaluation runs; RLC uses four checkpoints and public leaderboard results. These resource differences accompany comparisons.

Complete original PDF block S5.T4
Table 4: Overall results on the 2025 BEHAVIOR Challenge (50 tasks, 10 instances each). Task Success Score is the challenge ranking metric (task progress). The first place solution by the Robot Learning Collective (RLC) [58] uses a set of 4 checkpoints; π0.5\pi_{0.5} [4] (4 epochs) and G0.5 each use a single checkpoint, averaged over two eval runs. Best/second best in bold/underline.
Original selectable table
Method Task Success Score (Ranking Metric) ↑\uparrow
RLC (1st place) [58] 0.2605
Comet (2nd place) [59] 0.1830
π0.5\pi_{0.5} (4 epochs)  [4] 0.2626
G0.5 (Ours, 1 epoch) 0.2904
G0.5 (Ours, 4 epochs) 0.3136

5.3.2 Key Findings and Analysis

G0.5 scores 0.2904 after one epoch and 0.3136 after four, versus 0.2626 for pi0.5 at four epochs and 0.2605 for RLC. The relative pi0.5 gains are 10.6% and 19.4%, equivalent to 2.78 and 5.10 points on a 100 point scale. The headline 31.4% is the four-epoch result.

Matched downstream data supports transfer efficiency, but changes in architecture, foundation data and training prevent attributing the entire difference to a single component. More epochs improve the overall mean while several individual tasks regress.

Structured Action Decomposition Benefits Mobile Manipulation.

Factoring navigation and arm groups is a plausible reason for strong mobile manipulation. Examples include moving boxes,+0.35 over pi0.5, trash,+0.30, and loading a car,+0.28. No flat-tokenization ablation in this report isolates grouping as their cause.

Pre-Training Distribution Shapes Downstream Strengths.

Frequent pick/place training aligns with strengths such as mousetraps and gift baskets. Microwave popcorn remains 0.55 versus pi0.5’s 0.95; hot dogs improves 0.45→0.90 with more fine-tuning. The reported 29 wins/15 losses/6 ties across tasks describe this comparison, not universal task dominance.

Visual Memory Pre-Training Improves Long-Horizon Performance.

Temporal pretraining is proposed to improve single-frame representations of moving-camera scenes. Because there is no otherwise-matched no-memory pretraining ablation, this remains an interpretation of task patterns. Dropping final history tokens and using one-frame deployment also limits what can be inferred about long-term memory.

5.4 Real-World Fine-Tuning Evaluation

Four semantic tasks create six task-embodiment settings on R1-Lite and R1-Pro. Shared tasks test adaptation across these platforms after fine-tuning; this is not an unseen-body zero-shot experiment.

Robot Embodiments.

R1-Lite has two 6 DoF arms, 3 DoF torso and mobile base; R1-Pro has two 7 DoF arms, 4 DoF torso and mobile base. A common output representation supports both after source-specific preprocessing and task adaptation.

Baselines and Fine-Tuning Protocol.

G0.5, pi0.5andGR00T-N1.7 use the same task data and sixteenH20GPUs for the same 4–10 hour duration per setting, with aligned observations/actions/inference/control. Equal wall time is a practical budget match, not equal gradient updates or identical pretraining resources.

Evaluation Protocol.

Fifteen real episodes per setting are interleaved across models. Full-task success is separate from a stage-based process score. The text promises detailed stage scoring in the appendix, but this 32 page version’s appendix contains only BEHAVIOR/RoboTwin tables; process-score maxima and complete rubrics are not supplied.

Observation Setup.

Both robots use two wrist RGB cameras and an external global camera. This differs from the DROID two-view setup and conditions physical comparisons.

5.4.1 Task Definitions and Evaluation Metrics

Towel folding grasps, unfolds, flattens, folds and places a deformable towel; carton folding assembles a flat carton. Both are tested on both bodies. R1-Lite additionally unzips, fills and closes a pencil case; R1-Pro transfers five boxes and aligns grooves while stacking. Figure 7 retains all six scenes.

These tasks require sequential contact and bimanual coordination. Their process scores should not be normalized or compared across task rubrics whose full definitions are absent.

Complete original PDF block S5.F7.sf1
(a) R1-Lite: Folding Towel
Complete original PDF block S5.F7.sf2
(b) R1-Lite: Folding Carton
Complete original PDF block S5.F7.sf3
(c) R1-Lite: Pencil-Case Packing
Complete original PDF block S5.F7.sf4
(d) R1-Pro: Folding Towel
Complete original PDF block S5.F7.sf5
(e) R1-Pro: Folding Carton
Complete original PDF block S5.F7.sf6
(f) R1-Pro: Box Transfer and Stacking
Figure 7: Overview of real-world fine-tuning evaluation settings. We evaluate four manipulation tasks instantiated as six task-embodiment settings across R1-Pro and R1-Lite. The R1-Pro settings include towel folding, carton folding, and box transfer and stacking, while the R1-Lite settings include towel folding, carton folding, and pencil-case packing. Towel folding and carton folding are shared across both embodiments, enabling cross-embodiment comparison under the same task objectives.
Complete original PDF block S5.F8.part1Complete original PDF block S5.F8.part2
Figure 8: Performance on real-world fine-tuning evaluation. We evaluate G0.5 against π0.5\pi_{0.5} and GR00T-N1.7 on four manipulation tasks instantiated as six task-embodiment settings across R1-Pro and R1-Lite. G0.5 achieves strong overall performance in both task success rate and process score, demonstrating robust long-horizon manipulation capability across different embodiments.

5.4.2 Quantitative Results

Across six settings, mean success is 76.7% G0.5, 53.3% pi0.5and 24.4% GR00T; process scores are 129.2, 105.2 and 68.9. G0.5 leads five settings, but pi0.5 leads R1-Pro stacking 93.3% versus 80.0%.

For the four shared towel/carton settings, success is 75.0% versus 43.3% and 13.3%. G0.5’s R1-Pro/R1-Lite means are 75.6/77.8%. These are small-sample empirical averages without confidence intervals.

5.5 Pick-and-Place Benchmark

PP Bench distinguishes selecting the instructed object from completing its grasp and placement. This helps localize some failures, although its language-following metric checks approaching/attempting the object and does not independently verify every container or instruction constraint.

Dataset and Post-training Setup.

Fifty hours of R1-Lite demonstrations have 5–20 objects per scene and explicit object/container instructions. Nested 1/10/50 hour subsets train for 12/8/4 epochs. Increasing data therefore also changes total sample exposure and epoch count; it is not a constant-compute scaling experiment.

Evaluation Setting.

Each policy receives 64 paired real trials over 48 in-distribution objects and 16 categories absent from post-training, with 16 objects/containers arranged per scene. These unseen categories need not be absent from foundation training. Attempts toward the right object count for language following; successful grasp and correct-container placement count for task success.

Complete original PDF block S5.F9.part1
Figure 9: Pick-and-Place Benchmark setting. The evaluation set contains 64 object categories and 3 container categories. Each trial presents 16 randomly arranged objects and containers, requiring the robot to identify the instructed target among distractors and place it into the specified container. Each policy is evaluated across all object categories in real-world trials to reduce variance.

Results and Analysis.

G0.5 language-following/task-success pairs are 65.6/59.4 zero-shot, 62.5/57.8 at 1 hour, 71.9/65.6 at 10 hours and 84.4/75.0 at 50 hours. Thus 1 hour fine-tuning slightly reduces both metrics. At 50 hours, pi0.5scores 68.8/65.6, leaving 15.6 and 9.4 point gaps.

Additional box coordinates yield 85.9/76.6, a small numerical increase over 84.4/75.0 despite prose saying no improvement. Adding target object/container crops gives 98.4/84.4. Those auxiliary localization inputs change available information; gains do not demonstrate language-only grounding at 98.4%.

Complete original PDF block S5.F10
Figure 10: PP Bench evaluation results. Left: language following rate. Right: task success rate. We compare G0.5 and π0.5\pi_{0.5} across zero-shot, 1H, 10H, and 50H post-training settings.
Complete original PDF block S5.T5
Table 5: Ablation on referring context for PP Bench. We evaluate G0.5 with progressively richer referring context, i.e., auxiliary inputs appended to the standard instruction that help the policy identify the language-specified target object beyond its category name.
Original selectable table
Metric Name-only + Box Coord. + Target Visual
Language Following Rate 84.4% 85.9% 98.4%
Task Success Rate 75.0% 76.6% 84.4%

5.6 Zero-Shot Probe of CoT and Action Head

One pretrained checkpoint is switched between AR and optional FM heads and between CoT-on/off. Both heads condition on post-CoT states. PP uses 64 rollouts; Air Fryer and Cook Bacon each receive externally supplied instructions for five sequential stages and five rollouts per cell. No fine-tuning occurs within this probe.

Progress is defined as integer completed stages, 0–5. With five rollouts, a simple mean must fall in 0.2 increments, yet reported values include 2.1, 2.7 and 1.5. Additional pooling or fractional scoring is unspecified; preserve the reported averages without inventing a denominator.

Complete original PDF block S5.F11
Figure 11: Long-horizon zero-shot probe on Air Fryer and Cook Bacon (same pretrained G0.5 checkpoint with only inference-time switching). Progress score (0–5, number of completed sub-stages) and language-following rate across the four decoder ×\times CoT cells: both metrics show the same ordering, with AR++CoT clearly leading on each task.

Finding 1: CoT improves grounding and execution within long-horizon stage-conditioned rollouts.

CoT changes PP language following by about one trial: AR 65.6→67.2 andFM59.4→60.9. For long tasks, reported AR progress improves 2.4→3.8 AirFryer and 1.5→3.4 Bacon. This supports the tested stage-conditioned setup, not autonomous discovery or scheduling of all five stages.

Qualitative observation: instruction wording.

Adverbs, spatial qualifiers and alternative verbs sometimes changed successful rollouts in informal AR-only trials. The authors explicitly label this qualitative. No controlled wording sweep or statistical causal estimate is provided.

Finding 2: AR appears to follow the CoT more closely than the FM head.

FM progress improves 2.1→2.7 AirFryer and 1.2→2.0 Bacon, less than AR. CoT-on language following is 72 versus 48% and 64 versus 44%, respectively. Direct token attention versus pooled FM conditioning is proposed as an explanation, but not probed independently.

CoT correctness.

Human scoring finds roughly 90% correct CoT on PP, 85% AirFryer and 80% Bacon, with no systematic head difference. Separate rollouts diverge and therefore do not share identical CoT traces. Similar aggregate judged correctness does not fully control trace content or prove a decoding mechanism.

5.7 G0.5 Fine-Tunes with RL Out of the Box

One demonstration per LIBERO task initializes joint AR/FM post-training, then four tasks with comparable initial success are selected for GRPO. AR samples with temperature and exact token log-probabilities; FM uses an SDE/denoising-path approximation following RLinf.

Figure 12 reports 63.0→97.4% AR and 62.5→92.9% FM by 100 steps, with mean±standard-deviation bands. The selected four tasks, seed count, reward details and full RL hyperparameters are not enumerated here. Discrete code likelihood is exact for tokens, not automatically a continuous executed-action density when decoding is many-to-one.

Complete original PDF block S5.F12
Figure 12: GRPO fine-tuning: AR vs. FM. Average success rate over four LIBERO tasks during GRPO fine-tuning from a single demonstration per task (mean ±\pm std over seeds). Starting from AR and FM policies with comparable initial success rates, the AR policy converges faster, attains a higher final success rate, and shows lower run-to-run variance.

6 Conclusion

Results support autoregressive codes as a practical VLA interface with optional reasoning and RL fine-tuning. Component-level explanations remain hypotheses where controlled ablations are absent. Semi-transparent drawers, only seconds of history and lack of separate lower-body evaluation are acknowledged limits.

The conclusion’s broad claim that zero-shot PP language following exceeds the post-trained pi0.5 baseline needs a data-budget qualification:65.6% exceeds its 1/10 hour results, but not its 50 hour 68.8%. BEHAVIOR’s one-epoch gain is 0.2904, while the headline 0.3136 requires fourepochs.

7 Contributors

The PDF credits Galaxea Team, with Yicheng Liu as project lead and Hang Zhao as PI. Complete named contributor roles are retained below; catalog author metadata remains available for author-list membership.

Complete original PDF block source-contributors
Complete original contributor roles
Original selectable source block

Data engineering: Tao Jiang, Ke Dong, Xiaoshu Ren, Chenru Wu, Xiao Liu, Tianyuan Yuan, Zibin Dong, Zihan Guo, Zijie Zhao

Annotation & supplemental data: Tianyuan Yuan, Tao Jiang, Changxun Pan, Xinlei Zhang, Chenru Wu, Haonan Liu, Haodong Yang, Bowen Zhang

Policy training & research: Yicheng Liu, Zibin Dong, Tianyuan Yuan, Baijun Ye, Shicheng Cao, Shaoting Zhu, Xiao Liu

Policy evaluation & benchmarking: Anqi Yang, Zihan Guo, Baijun Ye, Zibin Dong, Tao Jiang, Yue Sun, Tianyuan Yuan, Tailai Cheng, Changxun Pan, Jianning Cui, Shicheng Cao, Haonan Liu, Shaoting Zhu, Zijie Zhao, Haoyu Zhang, Jiahui Niu, Shiduo Zhang

Writing & illustration: Yicheng Liu, Zibin Dong, Baijun Ye, Shicheng Cao, Haonan Liu, Tailai Cheng, Tao Jiang, Zihan Guo, Anqi Yang, Yue Sun, Tianyuan Yuan, Shiduo Zhang, Hang Zhao

Project lead: Yicheng Liu  • Project PI: Hang Zhao

Appendix A Appendix / supplemental material

The appendix contains all fifty BEHAVIOR task rows and all fifty RoboTwin Clean/Randomized rows. BEHAVIOR includes exact-zero tasks for every method and regressions after longer fine-tuning; RoboTwin’s hanging-mug 40/56 and turn-switch 74/80 show weaker cases hidden by a high mean.

These are the full supplemental contents of v 1. The real-world process-score rubric promised in the main text is absent; no missing scoring rules have been invented.

Complete original PDF block A1.T6
Table 6: Detailed results on the 2025 BEHAVIOR Challenge (50 tasks, 10 instances each). Task Success Score is the challenge ranking metric (task progress). The first place solution by the Robot Learning Collective (RLC) [58] uses a set of 4 checkpoints; π0.5\pi_{0.5} [4] (4 epochs) and G0.5 each use a single checkpoint, averaged over two eval runs. Best/second best in bold/underline.
Original selectable table
Task RLC [58] Comet [59] π0.5\pi_{0.5} (4 epochs) [4] G0.5 (1 epoch) G0.5 (4 epochs)
assembling gift baskets 0.2125 0.0000 0.2312 0.5188 0.4938
attach camera to tripod 0.0000 0.0000 0.0000 0.0000 0.0000
boxing books for storage 0.0000 0.0000 0.0000 0.0000 0.0000
bringing in wood 0.0667 0.5000 0.1667 0.3333 0.3500
bringing water 0.2667 0.9000 0.6500 0.8333 0.7000
can meat 0.0000 0.0000 0.0333 0.0111 0.0444
canning food 0.0100 0.0000 0.0550 0.0800 0.1100
carrying in groceries 0.1500 0.0000 0.1750 0.0750 0.1500
chop an onion 0.3000 0.0000 0.1750 0.4125 0.4000
chopping wood 0.1000 0.0000 0.0875 0.2375 0.2000
clean a patio 0.0000 0.0000 0.0000 0.0000 0.0000
clean a trumpet 0.0000 0.0000 0.0000 0.0000 0.0000
clean boxing gloves 0.2000 0.0000 0.0000 0.2250 0.1750
clean desk 0.1273 0.0000 0.2227 0.2864 0.2591
clean plates and food 0.1857 0.0000 0.3857 0.1786 0.1929
clear food to fridge 0.0800 0.0000 0.0800 0.1800 0.2700
collect childrens toys 0.4333 0.0000 0.4214 0.5857 0.5929
cook bacon 0.7571 0.0000 0.4214 0.0714 0.3214
cook cabbage 0.0000 0.0000 0.0500 0.1500 0.2000
cook hot dogs 0.8500 1.0000 0.9250 0.4500 0.9000
freeze pies 0.0143 0.1571 0.1357 0.0571 0.0429
get organized for work 0.0200 0.0000 0.0200 0.1050 0.1100
hanging pictures 0.0000 0.2000 0.0000 0.0000 0.0000
hiding Easter eggs 0.1444 0.2444 0.0778 0.0500 0.0500
loading the car 0.2000 0.0000 0.0000 0.2333 0.2833
make microwave popcorn 0.9000 0.7000 0.9500 0.1500 0.5500
make pizza 0.0000 0.0000 0.0000 0.0000 0.0000
move boxes to storage 0.6500 1.0000 0.2000 0.6250 0.5500
outfit basic toolbox 0.2571 0.1000 0.3429 0.1429 0.2500
pick up toys 0.3000 0.0000 0.1833 0.3167 0.4083
pick up trash 0.6667 0.7667 0.5500 0.8167 0.8500
prepare lunch box 0.5167 0.0000 0.5417 0.5667 0.5667
put away Halloween decor. 0.2000 0.5000 0.3714 0.4857 0.5286
put dishes away 0.2714 0.0000 0.5250 0.1393 0.2464
put shoes on rack 0.5000 0.5400 0.3650 0.6450 0.5650
put up Christmas decor. 0.4333 0.0000 0.5611 0.4667 0.3889
rearrange kitchen furn. 0.3000 0.3750 0.3875 0.3500 0.3625
set up coffee station 0.1500 0.2167 0.2833 0.2500 0.1917
setting mousetraps 0.3333 0.0000 0.1083 0.5083 0.5667
setting the fire 0.3250 0.2000 0.0750 0.3125 0.1250
slicing vegetables 0.1889 0.0000 0.2278 0.4444 0.2611
sorting household items 0.0625 0.0000 0.1938 0.1437 0.2062
sorting vegetables 0.4769 0.0000 0.6000 0.5231 0.6231
spraying for bugs 0.2500 0.1000 0.1000 0.2000 0.1500
spraying fruit trees 0.3000 0.3500 0.2000 0.1250 0.1500
storing food 0.3750 0.0000 0.5750 0.4938 0.6625
tidying bedroom 0.4000 0.0000 0.3500 0.5667 0.6167
turning on radio 0.6000 1.0000 0.1500 0.3000 0.0500
wash a baseball cap 0.4500 0.3000 0.6000 0.6500 0.7500
wash dog toys 0.0000 0.0000 0.3750 0.2250 0.2167
Overall 0.2605 0.1830 0.2626 0.2904 0.3136
Complete original PDF block A1.T7
Table 7: Per-task success rates of G0.5 on RoboTwin 2.0 under clean and randomized evaluation settings.
Original selectable table
Task Clean Rand.
Adjust Bottle 100 100
Beat Block Hammer 100 96
Blocks Ranking RGB 100 100
Blocks Ranking Size 96 96
Click Alarmclock 100 100
Click Bell 100 100
Dump Bin Bigbin 94 96
Grab Roller 100 100
Handover Block 98 84
Handover Mic 98 100
Hanging Mug 40 56
Lift Pot 100 100
Move Can Pot 98 94
Move Pillbottle Pad 96 96
Move Playingcard Away 100 98
Move Stapler Pad 80 70
Open Laptop 100 98
Open Microwave 92 88
Pick Diverse Bottles 84 86
Pick Dual Bottles 96 88
Place A2B Left 94 94
Place A2B Right 94 90
Place Bread Basket 98 98
Place Bread Skillet 96 90
Place Burger Fries 100 98
Place Can Basket 68 72
Place Cans Plasticbox 96 100
Place Container Plate 100 98
Place Dual Shoes 88 94
Place Empty Cup 100 100
Place Fan 100 94
Place Mouse Pad 84 84
Place Object Basket 90 92
Place Object Scale 96 94
Place Object Stand 100 96
Place Phone Stand 94 98
Place Shoe 98 94
Press Stapler 92 92
Put Bottles Dustbin 90 90
Put Object Cabinet 90 90
Rotate QRcode 96 98
Scan Object 100 92
Shake Bottle 100 100
Shake Bottle Horizontally 100 100
Stack Blocks Three 98 100
Stack Blocks Two 100 100
Stack Bowls Three 94 86
Stack Bowls Two 96 90
Stamp Seal 88 92
Turn Switch 74 80
Average 93.72 92.84
Concept demonstration

Open image ↗