Organizations: Sun Yat-sen University · X-Era AI Lab · Guangdong University of Technology · Guangdong Key Laboratory of Big Data Analysis and Processing
Vision-Language-Action (VLA) models offer a unified framework for robotic manipulation by integrating visual perception, language understanding, and control generation. However, existing VLA systems still struggle to generalize across diverse tasks, scenes, and camera viewpoints, and often produce coarse or unstable actions. We argue that these limitations are closely tied to the structural properties of actions in VLA settings, including the inherent multi-peaked nature of action distributions, the token-based symbolic reasoning of pretrained VLM/VLA backbones, and the effective finite resolution imposed by real-world robotic control. Motivated by these properties, we introduce E0, a tweedie discrete diffusion framework that formulates action generation as iterative denoising over quantized action tokens. By operating in a discrete action space with a principled diffusion process, E0 naturally aligns with token-based reasoning, supports fine-grained yet executable action control, and avoids the distributional mismatch of masking-based discrete diffusion. Experiments on LIBERO, VLABench, ManiSkill, and a real-world Franka arm demonstrate that E0 achieves state-of-the-art performance across diverse environments, outperforming strong baselines by 8.1% on average.
Figures & tables
Figure 1: Overview of action modeling paradigms. (a) Discrete modeling : Autoregressive (AR) approaches ( Brohan et al., 2022 ; Zitkovich et al., 2023 ; Kim et al., 2024 ) and recent mask-based discrete diffusion methods ( Liang et al., 2026 ; Song et al., 2026 ) , which operate over a small discrete action vocabulary. (b) Continuous modeling : Continuous diffusion–based policies ( Liu et al., 2024 ; Xu et al., 2025 ) and AR–diffusion hybrids ( Black et al., 2024 ; Intelligence et al., 2025 ; Lin et al., 2025 ) that regress continuous actions. (c) Our approach : E0 integrates AR-style conditioning with Tweedie discrete diffusion, preserving compatibility with pretrained vision–language backbones and allowing fine-grained action generation.
Figure 2: Architectural illustration of E0. (a) Overview of E0. (b) Training and inference pipeline, showing how inputs are encoded, diffused, and decoded into executable action sequences.
Model
LIBERO ( Liu et al., 2023 )
ManiSkill ( Tao et al., 2024 )
VLABench ( Zhang et al., 2024 )
Avg.
Spatial
Object
Goal
Long
Insert
Pick
Stack
Plug
Push
Toy
Fruit
Painting
Poker
Mahjong
Continuous Diffusion
Diffusion Policy ( Chi et al., 2025 )
78.3
92.5
68.3
50.5
0.0
40.0
80.0
0.0
88.0
-
-
-
-
-
-
MDT ( Reuss et al., 2024 )
78.5
87.5
73.5
64.8
-
-
-
-
-
-
-
-
-
-
-
RDT ( Liu et al., 2024 )
-
-
-
-
13.2
77.2
74.0
1.2
100.0
-
-
-
-
-
-
Dita ( Hou et al., 2025 )
84.2
96.3
85.4
63.8
-
-
-
-
-
-
-
-
-
-
-
Table 1: Performance comparison across multiple benchmarks (LIBERO, VLABench, ManiSkill). Each benchmark contains several task subsets, and the final column reports the average success rate (%) . A value of 0 indicates the model completely failed on this task, while “–” denotes cases where evaluation was not conducted due to model or dataset constraints. A “–” in the final average column means the model was not tested on all benchmarks (direct averaging would be misleading). Bold and underline numbers indicate the best and second-best results per column.
Figure 3: Benchmarks for evaluation. (a) LIBERO ( Liu et al., 2023 ) : tasks with varying objects, layouts, and goals, including long-horizon settings. (b) ManiSkill ( Tao et al., 2024 ) : diverse fine-grained manipulation skills (push, pick, stack, insert, plug). (c) VLABench ( Zhang et al., 2024 ) : open-ended tasks requiring language grounding and commonsense reasoning (select toy/fruit/painting/poker/mahjong).
Figure 4: Comparison on the VLABench benchmark. In the task “pick up the spade 3” , our E0 correctly identifies and precisely grasps the target card, showing superior multimodal reasoning and control ability.
Figure 5: Performance on real-world robotic experiments. (a) Short-horizon tasks ( press button , close door , pull drawer , pick block , stack block , stack block unseen ). (b) Long-horizon tasks ( pick block twice , pull drawer and put in block , and put in plate and close door ).
Model
Short-horizon Tasks
Long-horizon Tasks
Average
Pick
Press
Stack
Pull
Close
Pick Twice
Open & Put
Put & Close
SR (%) ↑
SR (%) ↑
SR (%) ↑
SR (%) ↑
SR (%) ↑
SR (%) ↑
SR (%) ↑
SR (%) ↑
SR (%) ↑
π0 FAST Pertsch et al. (2025)
20.0
20.0
5.0
10.0
15.0
10.0
0.0
0.0
10.0
π0 Black et al. (2024)
60.0
60.0
40.0
30.0
45.0
45.0
55.0
10.0
43.1
π0.5 Intelligence et al. (2025)
65.0
47.5
55.0
42.5
40.0
47.5
50.0
15.0
45.3
E0 (ours)
75.0
70.0
52.5
32.5
35.0
60.0
40.0
20.0
48.1
Table 2: Task success rates in real-world robot experiments. We evaluate policy performance across both short-horizon and long-horizon manipulation tasks. For the π0.5 and E0, we conduct an additional evaluation run and report the average results across the two runs. Bold numbers indicate the best performance in each column.
Figure 6: Comprehensive ablation analysis of key hyperparameters in the LIBERO environments. We investigate four crucial design factors influencing our model: (a) Discretization bins —increasing bin granularity enhances precision up to 2048 bins, beyond which gains saturate; (b) Action horizon —a moderate prediction length balances reactivity and temporal consistency; (c) Action dimensions —configured action-slot width, where the smallest sufficient width achieves the highest average success rate without redundant padded positions; and (d) One-hot smooth factor —moderate decay values smooth discrete logits, stabilizing diffusion and improving overall success rate.
Figure 7: Comparison on the VLABench Select Painting task. E0 shows more consistent visual–language grounding and smoother control than π0.5 .
Appendix figures & tables29 assets
Supplementary material from the paper’s appendix.
Appendix
Model
Spatial
Object
Goal
Long
Hz
SR (%) ↑
n
t
n
t
n
t
n
t
π0 Black et al. [2024]
3024
226.426
3827
278.521
3302
239.643
7620
556.740
13.659
94.15
π0.5 Intelligence et al. [2025]
5769
374.863
7232
452.431
6018
378.104
14403
902.264
15.858
96.65
E0 (ours)
5457
474.433
7064
597.437
5876
497.646
13641
1141.340
11.819
97.65
Appendix
Table 3: Efficiency and inference statistics on LIBERO tasks. We report the number of inferences ( n ), total inference time ( t , in seconds), average control frequency (Hz), and overall success rate (SR).
bins
Libero-Spatial
Libero-Object
Libero-Goal
Libero-Long
Average
SR (%) ↑
SR (%) ↑
SR (%) ↑
SR (%) ↑
SR (%) ↑
128
0.0
0.0
0.0
0.0
0.00
256
87.2
93.8
75.8
45.4
75.55
512
87.8
93.0
80.2
58.8
79.95
1024
86.2
95.6
84.6
60.2
81.65
2048
90.8
95.8
84.4
66.2
84.30
Appendix
Table 4: Ablation study on discretization bins in the LIBERO environments. We evaluate different numbers of discretization bins used to discretize continuous action values. Unlike autoregressive VLAs that typically fix the bin size to 256, our model supports arbitrary discretization granularity. Performance improves with higher resolution and peaks at 2048 bins, indicating that this level of precision is sufficient for accurate action representation. Bold numbers mark the best performance per column.
dims
Libero-Spatial
Libero-Object
Libero-Goal
Libero-Long
Average
SR (%) ↑
SR (%) ↑
SR (%) ↑
SR (%) ↑
SR (%) ↑
8
90.8
95.8
84.4
66.2
84.30
16
86.8
93.4
80.8
62.6
80.90
32
85.8
96.4
79.4
58.4
80.00
Appendix
Table 5: Effect of configured action-slot width on LIBERO. The physical action dimensions remain fixed, while the model’s internal action-slot width Dslot varies. Larger configurations introduce zero-padded positions without changing the executed control channels. Bold numbers mark the best performance per column.
decay
Libero-Spatial
Libero-Object
Libero-Goal
Libero-Long
Average
SR (%) ↑
SR (%) ↑
SR (%) ↑
SR (%) ↑
SR (%) ↑
0.1
90.8
95.8
84.4
66.2
84.30
0.5
93.2
95.8
83.6
63.2
83.95
1.0
89.6
95.4
84.2
62.4
82.90
Appendix
Table 6: Ablation study on one-hot decay in the LIBERO environments. We examine the influence of the decay coefficient applied to one-hot action vectors prior to noise addition. This decay attenuates the sharp peaks of discrete logits, reducing the dominance of high-confidence tokens during diffusion. Smaller decay values lead to improved stability and higher overall success rates. Bold numbers mark the best performance per column.
Action Horizon
Libero-Spatial
Libero-Object
Libero-Goal
Libero-Long
Average
SR (%) ↑
SR (%) ↑
SR (%) ↑
SR (%) ↑
SR (%) ↑
1
74.8
58.0
60.4
11.2
51.10
5
84.6
80.8
72.2
28.0
66.40
10
86.0
94.6
75.6
47.0
75.80
20
78.4
95.2
79.2
52.2
76.25
30
71.4
81.4
69.8
41.6
66.05
Appendix
Table 7: Ablation study on the number of action horizons in the LIBERO environments. We analyze how the length of the predicted action chunk affects the success rate across four tasks. A moderate horizon yields the highest average success rate, while both very short and very long horizons lead to degraded performance, indicating a trade-off between reactivity and temporal consistency. Bold numbers mark the best performance per column.
Normalization Type
Libero-Spatial
Libero-Object
Libero-Goal
Libero-Long
Average
SR (%) ↑
SR (%) ↑
SR (%) ↑
SR (%) ↑
SR (%) ↑
Mean–Std
13.2
0.0
9.2
7.8
7.6
Quantile
90.8
95.8
84.4
66.2
84.3
Appendix
Table 8: Ablation study on normalization statistics in LIBERO environments . Comparison between mean–std normalization and quantile-based normalization. Bold numbers mark the best performance per column.
Figure 8: Comparison on the LIBERO benchmark. In the task pick up the milk and place it in the basket , our E0 successfully grasped the object with the correct posture and completed the task without knocking it over.
Model
Libero-Spatial
Libero-Object
Libero-Goal
Libero-Long
Average
SR (%) ↑
SR (%) ↑
SR (%) ↑
SR (%) ↑
SR (%) ↑
Diffusion Policy Chi et al. [2025]
78.3
92.5
68.3
50.5
72.4
MDT Reuss et al. [2024]
78.5
87.5
73.5
64.8
76.1
OpenVLA Kim et al. [2024]
84.7
88.4
79.2
53.7
76.5
Octo Mees et al. [2024]
78.9
85.7
84.6
51.1
75.1
Dita Hou et al. [2025]
84.2
96.3
85.4
63.8
82.4
Appendix
Table 9: LIBERO task performance (success rate, SR) results. Evaluation across four subsets— Spatial , Object , Goal , and Long-horizon . Diffusion-based and autoregressive VLAs are compared, showing that large-scale generalist models significantly improve robustness and our proposed E0 achieves the highest average SR, demonstrating strong cross-task generalization and stability. Furthermore, we evaluated E0 using three independent runs under identical settings. Bold numbers denote the best results per column.
Figure 9: Comparison on the ManiSkill benchmark. In the task PegInsertionSide , our E0 achieved the best performance in this highly dexterous task. It was able to precisely align the peg with the hole and successfully insert it, while other models performed poorly.
Model
PegInsertionSide
PickCube
StackCube
PlugCharger
PushCube
Average
SR (%) ↑
SR (%) ↑
SR (%) ↑
SR (%) ↑
SR (%) ↑
SR (%) ↑
Diffusion Policy Chi et al. [2025]
0.0
40.0
80.0
0.0
88.0
41.6
OpenVLA Kim et al. [2024]
0.0
8.0
8.0
0.0
8.0
4.8
Octo Mees et al. [2024]
0.0
0.0
0.0
0.0
0.0
0.0
DD-VLA Liang et al. [2026]
0.0
4.0
70.0
0.0
100.0
34.8
RDT Liu et al. [2024]
13.2
77.2
74.0
1.2
100.0
53.1
Appendix
Table 10: ManiSkill task performance (success rate, SR) results. Evaluation on five challenging manipulation tasks— PegInsertionSide , PickCube , StackCube , PlugCharger , and PushCube —from the ManiSkill benchmark. Compared with both diffusion-based and autoregressive VLAs, our proposed E0 achieves the highest average SR. While prior models often fail on precise insertion or long-horizon coordination, E0 maintains stable performance, highlighting the effectiveness of discrete diffusion-based action representation. Bold numbers denote the best results per column.
Target location
Trials
Successes
SR (%)
Left
9
0
0.0
Center
31
18
58.1
Right
10
0
0.0
Overall
50
18
36.0
Appendix
Table 11: Target-position breakdown on Select Painting. We report π0.5 ’s task outcomes over 50 evaluation rollouts, grouped by the ground-truth target location. Success rate is computed within each row. Center-directed behavior was observed in all 50 rollouts, but this observation does not imply that the center button was successfully pressed in every trial.
Figure 10: Additional qualitative comparisons on the VLABench select painting task. This figure extends the results presented in Figure 7 , showcasing additional examples and outcomes.
Model
Select Toy
Select Fruit
Select Painting
Select Poker
Select Mahjong
Average
SR (%) ↑
SR (%) ↑
SR (%) ↑
SR (%) ↑
SR (%) ↑
SR (%) ↑
OpenVLA Kim et al. [2024]
0.0
0.0
20.0
0.0
4.0
4.80
DD-VLA Liang et al. [2026]
24.0
36.0
0.0
12.0
0.0
14.40
π0 Black et al. [2024]
54.0
48.0
16.0
6.0
7.0
26.20
π0 FAST Pertsch et al. [2025]
46.0
42.0
26.0
30.0
20.8
32.97
π0.5 Intelligence et al. [2025]
24.0
18.0
36.0
20.0
6.5
20.90
Appendix
Table 12: VLABench task performance (success rate, SR) results. Evaluation on five language-conditioned visual reasoning tasks from VLABench, including Select Toy , Select Fruit , Select Painting , Select Poker , and Select Mahjong . These tasks require fine-grained visual grounding and action understanding under compositional instructions. Compared to baseline models, our proposed E0 achieves the highest average success rate, showing stronger instruction following and cross-domain transferability. Bold numbers denote the best results in each column.
Model
Select Toy
Select Fruit
Select Painting
Select Poker
Select Mahjong
Average
PS ↑
PS ↑
PS ↑
PS ↑
PS ↑
PS ↑
OpenVLA Kim et al. [2024]
0.00
0.00
0.20
0.1066
0.1800
0.0973
π0 Black et al. [2024]
0.76
0.72
0.16
0.1000
0.0814
0.3643
π0 FAST Pertsch et al. [2025]
0.72
0.68
0.26
0.4200
0.3333
0.4827
π0.5 Intelligence et al. [2025]
0.52
0.49
0.36
0.2867
0.1739
0.3661
E0 (ours)
0.76
0.65
0.12
0.7733
0.2500
0.5107
Appendix
Table 13: VLABench task performance (process score, PS) results. We report process-level performance across five language-conditioned selection tasks from VLABench Zhang et al. [2024] . While baselines such as π0 Black et al. [2024] , π0 FAST Pertsch et al. [2025] , and π0.5 Intelligence et al. [2025] demonstrate task-specific strengths, our proposed E0 attains the highest average process score, reflecting more consistent and fine-grained reasoning throughout multi-step action modeling. Bold numbers indicate the best performance in each column.
Figure 11: Comparison on the RoboTwin benchmark. In the single-arm tasks, our E0 shows the performance deviations from π0 across four tasks.
Figure 12: Comparison on the RoboTwin benchmark. In the dual-arm tasks, our E0 shows the performance deviations from π0 across four tasks.
Task (Left)
π0 SR (%)
E0 SR (%)
Diff
Task (Right)
π0 SR (%)
E0 SR (%)
Diff
Single-Arm Tasks
stamp seal
27.0
40.0
+13.0
adjust bottle
41.0
97.0
+56.0
turn switch
39.0
20.0
-19.0
beat block hammer
66.0
73.0
+7.0
Dual-Arm Tasks
click alarmclock
44.0
37.0
-7.0
blocks ranking rgb
17.0
34.0
+17.0
click bell
28.0
7.0
-21.0
blocks ranking size
7.0
11.0
+4.0
dump bin bigbin
58.0
54.0
-4.0
grab roller
86.0
94.0
+8.0
Appendix
Table 14: RoboTwin results across all 50 tasks. Red numbers indicate improvement, while green numbers indicate decline. Compared comprehensively with π0 , our model E0 achieves the highest average success rate across single-arm, dual-arm, and overall task categories.
Figure 13: Real-world evaluation setup . Observations are captured using a side-view camera and a wrist-mounted camera, while an additional tripod-mounted camera records the scene from a rear-side perspective.
Figure 14: Comparison of keyframes in the real-world pick twice task under unseen scenarios. During evaluation, we tested the model across various unseen scenarios. In most cases, the model successfully completed the task. In a few cases, task quality is slightly compromised but still acceptable—for example, placement deviations caused by an oversized green dish, or changes in the order of object picking due to color variations.
Figure 15: Comparison of keyframes in the real-world pick twice task with and without human intervention. We introduce human perturbation by manually shifting the target cube to disrupt the model’s original plan. After the interruption, the model is able to promptly adapt and replan its actions.
Figure 16: Comparison of keyframes in real-world short-horizon tasks.
Figure 17: Qualitative results on the real-world Pick Vegetables Twice task under randomly arranged tabletop scenes . Each row shows one complete execution sequence consisting of two consecutive pick-and-place operations. Top row: carrot and cucumber. Middle row: potato and eggplant. Bottom row: pepper and corn. Across all settings, the surrounding distractor objects are placed in completely random configurations, demonstrating the robustness of our policy under visually diverse and cluttered real-world environments.
Figure 18: RoboTwin single-arm tasks. A total of 27 diverse single-arm manipulation tasks are included, covering a wide range of objects and actions.
Figure 19: RoboTwin dual-arm tasks. A total of 23 dual-arm manipulation tasks are included, covering a wide range of objects and actions. Some tasks require coordinated control between both arms, posing higher challenges for the model.
Model
Libero-Spatial
Libero-Object
Libero-Goal
Libero-Long
Average
SR (%) ↑
SR (%) ↑
SR (%) ↑
SR (%) ↑
SR (%) ↑
π0 Black et al. [2024]
28.0
16.6
31.6
2.6
19.7
π0 Black et al. [2024] + view
53.8
75.2
57.8
16.2
50.8 (+31.1)
E0 (Ours)
74.4
85.2
77.4
28.8
66.5
E0 (Ours) + view
92.0
96.8
83.0
63.8
83.9 (+17.4)
Appendix
Table 15: Evaluation under camera viewpoint perturbations in the LIBERO environments. We evaluate the robustness of different models when camera positions and orientations are dynamically perturbed.
Multimodality
Number of Queries
Mean-to-Nearest-Mode Error ↓
Low
3,333
0.3731
Medium
3,334
0.4307
High
3,333
0.5458
Appendix
Table 16: Effect of action multimodality on mean-action deviation. We report the normalized distance between the conditional mean action and its nearest empirical action mode on LIBERO demonstrations. Lower values indicate smaller geometric deviation.
Model
Low ↓
Medium ↓
High ↓
π0.5
0.7446
0.7834
0.8214
E0 (ours)
0.6980
0.7299
0.7705
Appendix
Table 17: Nearest-mode prediction error across multimodality levels. E0 and π0.5 are evaluated on the same LIBERO observations. Lower values indicate greater proximity to empirical action modes.
Metric
π0.5
E0 (ours)
Improvement
Nearest-mode error ↓
0.7831
0.7328
6.43% lower
Off-mode rate ↓
11.07%
8.03%
27.41% lower
Target-action ADE ↓
0.8286
0.5680
31.44% lower
Mode selection ↑
78.50%
87.63%
+9.13 pp
Appendix
Table 18: Mode consistency and action prediction accuracy on LIBERO. Results are computed over 3,000 paired observations. Relative improvements are reported for error metrics, while mode-selection improvement is reported in percentage points.
CE
Categorical F-R proxy
Scaled score proxy
Step
Value
Relative
Value
Relative
Value
Relative
10
1.013256
1.0000
9.606386
1.0000
5.433×10−3
1.0000
5,000
0.095172
0.0939
1.299624
0.1353
2.995×10−4
0.0551
10,000
0.080838
0.0798
1.184651
0.1233
3.767×10−4
0.0693
15,000
0.062705
0.0619
1.007325
0.1049
2.552×10−4
0.0470
20,000
0.047389
0.0468
0.839231
0.0874
2.089×10−4
0.0384
Appendix
Table 19: Training diagnostics at selected optimization steps. We report cross-entropy, a categorical Fisher–Rao proxy, and a time-scaled score-error proxy. Relative values are normalized by the corresponding measurement at step 10. Scientific-notation entries are rounded for display.
Vision-Language-Action (VLA) models have established a powerful paradigm for generalist robotic manipulation by grounding control into the semantic reasoning of VLMs. Prevailing architectures typically model actions continuously via diffusion or flow processes, or discretely through either autoregressive generation or parallel decoding. Recently, Discrete Diffusion VLAs (dVLAs) have emerged as a distinct alternative, unifying vision, language, and action into a single discrete token space via masked generative modeling. While combining iterative refinement with unified representations, its training has thus far been restricted to Supervised Fine-Tuning (SFT), leaving the potential of Reinforcement Learning (RL) for further policy refinement largely unexplored. A fundamental challenge in RL for dVLAs is that the marginal probability of the final action generated by dVLAs remains intractable. To solve this problem, we propose dVLA-RL, shifting the learning objective from the marginal action probability to the joint probability of the sampled generation path. Specifically, by modeling the denoising process as a Markov Decision Process (MDP), we mathematically formulate this path probability as a product of step-wise transitions. This trajectory-level objective provides a unified formulation that natively accommodates variable denoising steps. Leveraging this intrinsic fexibility, we introduce a unified step scheduling approach for complex multi-task learning, tailoring denoising steps to specific task complexities to maximize both success rates and computational effciency. Extensive evaluations demonstrate that our approach achieves a success rate of 99.7% on LIBERO. Furthermore, it establishes strong VLA-based results on RoboTwin 2.0 by delivering a 30.6% improvement over the SFT baseline, remaining competitive with strong World-Action Model baselines.
Yuhao Wu, Yitian Liu, Weijie Shen +13
ScaleLab at SJTU · 1Shanghai Jiao Tong University · 3Baidu AI Cloud +3
Most vision-language-action (VLA) models map observations directly to actions without explicit intermediate planning, which limits performance on long-horizon tasks where early mistakes compound. We propose Coarse-to-Control, a plan-execute VLA that introduces planning natively in the action-token space. The key idea is to let the policy first predict a compact sequence of coarse action tokens that summarize the intended future trajectory, and then generate executable action tokens conditioned on this plan. Because both planning and execution share a unified discrete action vocabulary, the plan stays close to the control manifold and provides directly actionable guidance rather than an abstract hint that must be translated back to motor commands. Experiments on LIBERO, SimplerEnv-WidowX, and real-world manipulation tasks show that action-token planning consistently improves over direct action generation, with the largest gains on long-horizon multi-stage tasks.
Jinhao Wu, Shiduo Zhang, Yicheng Liu +9
1Nanjing University · 2Shanghai Innovation Institute · 3Fudan University +1
Diffusion-based vision-language-action (VLA) models cascade pretrained vision-language models (VLMs) with diffusion action experts, but typically restrict the VLM to providing static conditioning. This separation limits how much pretrained perception and reasoning can guide action generation, especially in out-of-distribution settings. We introduce JADE, a Joint Action DEnoiser that enables the VLM and action expert to jointly refine actions. Our key insight is that structured action latents allow the VLM to read the current noisy action together with visual observations and language instructions. At every denoising step, the VLM contextualizes the noisy action and guides the action expert through gated cross-attention, so that its guidance evolves with the action being generated. We further supervise the VLM to recover clean action latents from noisy ones, which gives it a direct action-learning signal and couples perception with control. Additionally, joint denoising enables JADE to generate actions with fewer sampling steps, helping offset the additional cost of VLM participation. An adaptive step-size schedule further balances task performance and inference latency. On RoboDojo, JADE achieves an average success rate of 12.67%, the highest among the evaluated π0.5-based methods, and is competitive with recent VLA and WAM foundation models pretrained at scale. It also improves performance on RoboTwin 2.0 and raises the average success rate across four real-world bimanual tasks by 20.8 percentage points compared with π0.5.
Xicheng Gong, Wenxuan Song, Chengze Liu +5
Peking University · The Hong Kong University of Science and Technology (Guangzhou) · Microsoft Research Asia