Public benchmark scores

Table 3 'Results on LIBERO'. Fine-tuned 100K steps (lr 1e-5) on the four suites; 10 tasks/suite, 50 demos/task; evaluated 50 rollout trials per task with benchmark initial states. Baselines: pi0.5 98.8/98.2/98.0/92.4/96.9, GR00T-N1.7 97.7/98.5/97.5/94.4/97.0, Xiaomi-Robotics-0 98.8/100.0/98.8/97.2/98.7.

Table 5 'Results on LIBERO'. Backbone Qwen-3.5 2B + 460M unified expert. Trials per task not stated in the extracted text (standard LIBERO protocol implied). Baselines: pi0.5 98.8/98.2/98.0/92.4/96.9, OpenVLA-OFT 97.6/98.4/97.9/94.5/97.1, Xiaomi-Robotics-0 98.8/100.0/98.8/97.2/98.7.

Qwen2.5-0.5B backbone, no robotic pre-training; trained per suite; each subtask repeated 50 times (500 trials per suite). Pro = updated policy implementation (Appendix); marked best (Bold*).

Fine-tuned per suite; printed with standard error and rank 1 in every column. Printed as 99.7 ± 0.2%.

Phase-II domain adaptation fine-tune on LIBERO (appendix config: Abs EEF, batch 128, 60K steps); trial count not stated in fetched text

Qwen2.5-0.5B backbone, no robotic pre-training; trained per suite; each subtask repeated 50 times (500 trials per suite).

Table 3 'Evaluations results on LIBERO. We compare our DexVLA with DP and OpenVLA' (Appendix). Columns are Spatial | Object | Goal | Average only - NO LIBERO-Long/LIBERO-10 column; 'Average' is over the three suites as printed. Baseline rows: DP 78.3/92.5/68.3/79.7, OpenVLA 84.7/88.4/79.2/84.1, pi0-FAST 96.4/96.8/88.6/93.9, pi0 96.8/98.8/95.8/97.1. DexVLA = Qwen2-VL 2B + 1B diffusion expert. Trials per task not stated in extracted text.

OpenVLA-OFT fine-tuned per suite; multi-input (third-person + wrist image + proprio); 500 trials per suite (10 tasks x 50 episodes); OFT paper by the OpenVLA authors

Checkpoint gs://openpi-assets/checkpoints/pi05_libero trained with the pi05_libero config; number of trials not stated. README undated; fetched 2026-09-18. The π0.5 paper (arXiv 2504.16054) prints no numeric benchmark tables

Fine-tuned per suite following OpenVLA-OFT protocol (failed demos and pause frames filtered; wrist + third-person views).

Fine-tuned per suite; 500 trials per suite (50 per task) averaged over three seeds. 'Full' = pretrained on the full latent-action pretraining mixture.

Table I 'Performance on Libero. We report the success rate for four LIBERO suites, for two category of models, with and without large scale action pretraining.' VLA-0 is in the 'without large-scale action pretraining' category. Base model Qwen2.5-VL-3B-Instruct; actions emitted as text (no vocabulary/architecture change); 50 episodes per task x 10 tasks per suite; eval uses action horizon 1 with ensemble of 8 predictions (README). Same-category baselines: pi0.5-KI 96.6/97.2/94.6/85.8/93.3, OpenVLA-OFT 94.3/95.2/91.7/86.5/91.9, SmolVLA 93.0/94.0/91.0/77.0/88.8. Numbers also on project page https://vla0.github.io/.

Fine-tuned by the OFT authors per suite; 500 trials per suite (10 tasks x 50 episodes). The π0 paper itself (arXiv 2410.24164) reports only bar-chart figures with no printed benchmark values

Pretrained only on Bridge-V2. Subsets: 95.2 / 95.4 / 91.9 / 87.5.

Larger VLM-backbone variant; X-VLA paper Table 2 reprints this row as 93.0/94.0/91.0/77.0/88.8

Pretrained only on action-free human video. Subsets: 91.2 / 94.2 / 90.2 / 79.4.

Fine-tuned per suite; 500 trials per suite. NORA-Long = action-horizon-5 variant.

Simulation fine-tuning 100k steps, batch 64; 10 trials per task, binary success; whether one multitask checkpoint or one per suite was not confirmed from the fetched text

Fine-tuned per suite by LoRA post-training on MolmoAct-7B-D (Table 7 hyperparameters).

Same protocol as π0 row

Smaller variant

AC = action chunking. Subsets: 85.6 / 89.4 / 80 / 63.

Fine-tuned per LIBERO suite (LoRA + spatial embedding adaption); success rate averaged over 3 random seeds with 500 trials per suite; printed with standard error. Printed as 88.2 ± 0.5%.

Fine-tuned separately per suite via LoRA r=32 on modified LIBERO data (no-op actions filtered, failed demos removed, 256px re-render, third-person camera only); 500 trials per suite averaged over 3 seeds (1500 trials); ± is std err

Octo fine-tuned per suite by OpenVLA authors on their modified LIBERO data; 500 trials x 3 seeds

Subsets: 85.6 / 87.8 / 77 / 45. 500 trials per suite.

Parameter-efficient adapters; README also lists Object 96%, Goal 94.4%, Long 83.2%, Simpler WidowX PEFT 66.7%; paper text (Sec. 5.2) quotes '93%' LIBERO and '54%' Simpler-WidowX for PEFT with 1% tunable params

Single checkpoint fine-tuned on all four suites with identical hyperparameters (max 20,000 steps, global batch 640, state dropout 0.2); 10 episodes per task (200 per suite); no average printed. README undated; fetched 2026-09-18

ZERO-SHOT and NON-STANDARD: LIBERO tasks rewritten as grasp-only 'pick up {object}' instructions (modifications in Tables 10-12); Spatial suite excluded; grasping any object of the target category counts as success; 50 randomized initial configurations per task. Baselines (fine-tuned): OpenVLA 33.7/56.6/65.4, pi0 62.7/79.4/93.8.

README value differs from paper Table 2 (80.4); VA in README = 76.4%. README undated; fetched 2026-09-18

Per-task: Pick Coke Can 95.3, Move Near 90.0, Open/Close Drawer 75.5, Open Top Drawer and Place Apple 62.0. Trained with spatial prompting (checkmark column).

Adapted to Simpler-Fractal (Google Robot). NOTE: the GitHub README reports 83.5% VM for checkpoint X-VLA-Google-Robot (different from paper)

Same eval protocol as Table 1

Per-task: Pick Coke Can 98.0, Move Near 83.8, Open/Close Drawer 71.3, Drawer Apple 52.8.

Fine-tuned on Google Fractal dataset. Per-task: 86.0 / 77.9 / 57.4.

Trained on OXE subsets, evaluated zero-shot in SIMPLER; per task Pick Coke Can 91.3, Move Near 85.0, Open/Close Drawer 71.8, Open Top Drawer & Place Apple 50.9; Table 7 confirms the default is DiT-Base (89M)

VA 51.3, WidowX VM 51.0 in same row

Zero-shot OXE-pretrained checkpoint. Per-task: Pick Coke Can 81.0, Move Near 69.6, Open/Close Drawer 59.3.

Fine-tuned on RT-1 dataset (MolmoAct-7B-D-Pretrain-RT-1). Per-task: 77.7 / 77.1 / 60.0.

Zero-shot OXE pre-trained checkpoint (MolmoAct-7B-D-Pretrain). Per-task: Pick Coke Can 71.3, Move Near 73.8, Open/Close Drawer 66.5.

As reprinted by X-VLA; VA 56.8, WidowX 27.8 in same row; X-VLA text says π0 achieves 55.7% on Simpler-WidowX in their own fine-tune (Table 3 area)

Zero-shot SIMPLER eval of the released OpenVLA checkpoint by the CogACT authors; 4-task average (Pick Coke Can 18.0, Move Near 56.3, Open/Close Drawer 63.0, Open Top Drawer & Place Apple 0.0)

4-task average (17.0 / 4.2 / 22.7 / 0.0). The SimplerEnv paper itself (arXiv 2405.05941, Table IV) prints only per-task fractions for Octo-Base VM (Pick Coke Can 0.170, Move Near 0.042, Open/Close Drawer 0.227, Open Top Drawer & Place Apple 0.000) with no average column. X-VLA paper Table 2 prints 16.8 for Octo VM (likely 3-task average)

NON-STANDARD task set: 6 Fractal tasks (pick_coke_can 100.0, pick_object 94.0, move_near 100.0, open_drawer 65.0, close_drawer 69.0, place_in_closed_drawer 7.0); visual-matching vs variant-aggregation not specified; fine-tuned 20,000 steps batch 1024; 10 episodes per task

Same non-standard 6-task Fractal set as N1.7 entry (95.0 / 87.0 / 81.0 / 0.0 / 44.0 / 5.0)

Extended Table 3(d) 'Simulation performance on SimplerEnv-Google Robot environments'. Values printed as fractions. Column structure: Pick Coke Can {Horizontal Laying, Vertical Laying, Standing, Average}, Move Near {Average}, Open/Close Drawer {Open, Close, Average}, Open Top Drawer and Place Apple {Average}. RoboVLMs row as printed: 1.000 | 0.910 | 1.000 | 0.970 | 0.888 | 0.565 | 0.537 | 0.551 | 0.325. No overall average printed. Row is identical to the 'In-domain Full Finetune' row of Extended Table 3(f) (fine-tuned on full RT-1/fractal data; tested tasks included). Caption does not name visual matching vs variant aggregation; the baseline rows (e.g. RT-1 Converged 0.857 / 0.442 / 0.730) coincide with SimplerEnv's published visual-matching values.

Zero-shot: pretrained Magma-8B 'without any domain-specific finetuning'. Caption: SE = SimplerEnv. Text: 8 tasks with 172 visual matching and variant aggregation scenarios; Table 2 does not label which setting; Fig. 8 caption says per-task results are averaged across visual matching and variant aggregation. Same column: OpenVLA 31.7, RT-1-X 34.2, Octo 6.0. Also confirmed on HF model card https://huggingface.co/microsoft/Magma-8B. No per-task table in the HTML; LIBERO (Fig. 10, 10 trajectories/task few-shot) and real WidowX (Fig. 9) results are figure-only.

Post-trained on BridgeData V2 for 80K steps (lr 3e-5), state-free policy input, then evaluated in SimplerEnv WidowX+Bridge without simulation-domain training. Per-task: Put Spoon on Towel 97.5, Put Carrot on Plate 75.0, Stack Green Block on Yellow Block 83.3, Put Eggplant in Yellow Basket 93.3. Baseline numbers compiled by the authors from prior studies (pi0.5 57.1, Xiaomi-Robotics-0 79.2). Trials per task not stated.

Per-task: Put Spoon on Towel 92.4, Put Carrot on Plate 85.4, Stack Green Block on Yellow Block 69.5, Put Eggplant in Yellow Basket 75.7. Baseline numbers compiled from prior studies (pi0.5 57.1, GR00T-N1.5 61.9, Xiaomi-Robotics-0 79.2). No Google Robot table in the paper. Training data / trials for this eval not stated in extracted text.

Per task: Spoon 4.2, Carrot 0.0, Stack 0.0, Eggplant 12.5. NOTE: X-VLA paper Table 2 prints 8.30 for OpenVLA on Simpler-WidowX (different eval run)

Per task 15.8 / 12.5 / 0.0 / 41.7; SimplerEnv paper Table V prints per-task success fractions 0.125 / 0.083 / 0.000 / 0.431 with no average. X-VLA Table 2 prints 23.4 for Octo WidowX

Per task 41.7 / 8.2 / 0.0 / 56.7; SimplerEnv paper Table V per-task fractions 0.472 / 0.097 / 0.042 / 0.569

Per task Put Spoon on Towel 71.7, Put Carrot on Plate 50.8, Stack Green Block 15.0, Put Eggplant in Basket 67.5

Same eval protocol as Table 2

Adapted on BridgeData; README lists the same 95.8% for X-VLA-WidowX; the paper text rounds to '96%'

As above

As above

NON-STANDARD task set: 7 Bridge tasks (spoon_on_towel 78.0, carrot_on_plate 58.0, put_eggplant_in_basket 53.0, stack_cube 48.0, put_eggplant_in_sink 2.0, close_drawer 97.0, open_drawer 100.0), not the standard 4-task WidowX suite; fine-tuned 20,000 steps batch 1024; 10 episodes per task, 300-step max; not directly comparable to other WidowX averages here

Same non-standard 7-task Bridge set as N1.7 entry (55.4 / 46.0 / 89.0 / 5.0 / 33.0 / 73.0 / 95.0); 10 episodes per task; N1.6 released 2025-12 (research page dated 15 Dec 2025); HF card and research page print no benchmark numbers

Zero-shot OXE-pretrained. Task success: Spoon on Towel 20.8, Carrot on Plate 20.8, Stack Green Block 25.0, Eggplant 70.8 (grasp rates also printed).

Fine-tuned on BridgeData V2. Task success: 16.7 / 25.0 / 29.2 / 100.0.

SimplerEnv WidowX (Bridge) tasks; overall task success. BridgeData is in the pretraining mix.

Only the action decoder trained for adaptation.

Per-task: Put Spoon on Towel 87.5, Put Carrot on Plate 67.9, Stack Green Block on Yellow Block 31.3, Put Eggplant in Yellow Basket 100.0.

Per-task: Put Spoon on Towel 63.6, Put Carrot on Plate 54.5, Stack Blocks 81.8, Put Eggplant in Basket 90.9.

NON-STANDARD protocol: pretrained on BridgeV2 without action labels, then fine-tuned on 100 SIMPLER trajectories; 24 rollouts per task over 4 Bridge tasks (stack G2Y 54.2, carrot2plate 45.8, spoon2towel 70.8, eggplant2basket 58.3). ActionVLA 63.5, OpenVLA 36.4.

NON-STANDARD protocol as Table 11 but pretrained on Something-Something V2 human video; 100 fine-tuning trajectories; 24 rollouts per task.

Extended Table 3(c) 'Simulation performance on SimplerEnv-WidowX+Bridge environments'. Values printed as fractions (0-1). Table reports both 'Grasp' sub-task rate and 'final' success per task; this entry is the 'final' column. No average column is printed. RoboVLMs row equals the 'Post Train' recipe in Extended Table 3(e) (OXE co-train, then post-train on Bridge V2; tested tasks excluded). Setting not labelled; SimplerEnv WidowX+Bridge tasks are visual-matching only. Grasp sub-task rates: spoon 0.583, carrot 0.458, block 0.375, eggplant 0.958.

Zero-shot pretrained Magma-8B, no fine-tuning. SimplerEnv Bridge/WidowX aggregate; setting not labelled in table (see Fig. 8 caption: averaged across visual matching and variant aggregation). Same column: OpenVLA 14.5, RT-1-X 1.1, Octo 15.9.

Blog table for SimplerEnv ('a few Simpler Env tasks'): MiniVLA + VQ h8 trained on BridgeData V2 (80% final action accuracy) vs OpenVLA (95%). Task labels as printed: Place carrot, Place spoon, Block stack, Pick eggplant (correspond to the SimplerEnv WidowX+Bridge tasks). OpenVLA row: 46 / 44 / 62 / 66. No average printed; rollout counts and visual-matching setting not stated. Blog notes MiniVLA 'performs worse' on eggplant.

Chain success 98.5 / 95.0 / 90.5 / 85.3 / 80.0 (%).

Adapted on CALVIN-ABC (Abs EEF, batch 128, 60K steps, ColorJitter); prior SOTA in table FLOWER 4.53; README lists 4.43 for X-VLA-Calvin-ABC_D

Chain success 99.1 / 94.6 / 88.8 / 82.8 / 76.5 (%).

Zero-shot to scene D. Per-step success (fractions): 0.980 / 0.936 / 0.854 / 0.778 / 0.704.

Ablation-study table; caption: 'All models are trained with reduced epochs for efficiency; scores are for relative comparison.' Chain success 95.4 / 88.2 / 82.2 / 78.2 / 72.1 (%).

CALVIN ABC->D; consecutive-task success 95.5 / 85.8 / 75.4 / 66.9 / 56.5 (%); third-view RGB only.

AMBIGUOUS CAPTION: caption says 'on RoboTwin' but the table header reads 'ABC -> D' with CALVIN-style 1/5..5/5 chain columns and Avg. Len., and Sec. IV-B discusses it alongside RoboFlamingo/OpenVLA CALVIN rows; treated here as CALVIN ABC->D. Per-step success (%): 97.4 / 86.2 / 73.4 / 64.6 / 53.4. The 0.5B row in the same table is 96.2 / 84.8 / 72.6 / 60.8 / 50.6, Avg. Len. 3.65 (a different configuration than Table VI's 3.68).

CALVIN ABC->D (unseen scene D). 1000 evaluation chains ('we test each method 1000 times'). Per-step success rates (%) as printed: 94.8 / 84.5 / 71.3 / 62.5 / 53.8. Backbone LLaVA-OneVision-0.5B; no large-scale robot-dataset pre-training; action chunk size 5; discrete action tokens (256 bins). Same table lists GR-1 3.06, OpenVLA 3.27, RoboFlamingo 2.47.

Zero-shot to scene D. Per-step success: 0.854 / 0.712 / 0.596 / 0.497 / 0.401. Baselines: HULC 0.67, RT-1 0.90.

RoboCasa 24-task kitchen suite, post-training with 30 demos per task; trial count not stated; page dated 11 June 2025

Baselines in Table 2: GR00T-N1 39.7, π0 37.8, π0+FAST 34.1

As reprinted by X-VLA

Table 2 'Results on RoboTwin 2.0. We report success rates under clean and randomized evaluation settings, together with their average.' Fine-tuned 4 epochs (lr 4e-5, batch 1024) on 2,500 clean + 25,000 randomized demos; success averaged over 100 trials per task per setting. Baselines: pi0.5 82.7/76.8/79.8, Fast-WAM 91.9/91.8/91.8, LingBot-VA 92.9/91.5/92.2. Per-task G0.5 rates in Table 7.

Table 3 'Results on RoboTwin. We report success rates under clean and randomized evaluation settings, together with their average.' (Main text calls the benchmark RoboTwin 2.0 and refers to this table as 'Table 4' - numbering inconsistency in the paper.) Baselines: pi0.5 82.7/76.8/79.8, InternVLA-A1 89.4/89.6/89.5, LingBot-VA 92.9/91.5/92.2, Fast-WAM 91.9/91.8/91.8. Trials per task not stated in extracted text.

Number as reprinted in X-VLA's comparison table; Hard = 13.7

Adapted on RoboTwin-2.0 (Abs EEF, batch 128, 60K steps, ColorJitter); README lists '70%' for X-VLA-RoboTwin2

As reprinted by X-VLA

Randomized Scenes 76.76%; average over 5 RoboTwin 2.0 tasks per LingBot README

Average success over 5 RoboTwin 2.0 tasks (per README); w/o depth 86.50; π0.5 82.74; same numbers in GitHub README

As above

Released 2026-07-08 per README; no paper on arXiv located; HF card and project page show charts without printed numbers

Paper labels the benchmark 'RoboTwin'; it cites RoboTwin 2.0 (ref [20]) and uses RoboTwin 2.0 task names. 8 tasks: Click Bell 81.0, Click Alarmclock 73.0, Lift Pot 39.0, Move Can Pot 28.0, Open Laptop 38.0, Pick Dual Bottles 26.0, Place Dual Shoes 8.0, Rotate Qrcode 29.0 (%). Baselines in same table: ACT 7.1, DP 34.5, RDT 48.9 (Seen). Trials per task not stated in extracted text. 'Seen' = clean setting; 'DR' = domain randomization.

2.25B variant avg 68.24; 0.24B variant avg 56.95

LeRobot SO100 arms; 60 trials per task (two arms x three evaluation settings x 10 trials); each number is the average SR across the three evaluation settings in Table 2. Baselines: GR00T N1.5 55.6, Pi0 70.4 (avg). SR@1 (first-attempt) 56.7.

AgileX Piper arm (7-DoF). Four tasks: Store screwdriver 93.3, Clean cutting board 100.0, Fold towel twice 46.7, Stack tower of hanoi 86.7; average score 2.63 (max 3). Trials per task not captured from the HTML.

Real WidowX-250, 9 tasks x 10 trials (multiple-object, OOD object, spatial categories); baselines RT-1 4.4, OpenVLA 40, SpatialVLA 11.1.

Real Franka Emika Panda; 3 tasks (knocking, covering, pick&place sink); 54 rollouts per model (18 per task) split across seen-obj/unseen-combo, unseen-obj, unseen-instruction. Columns: Seen Obj. Unseen Combo / Unseen Obj. / Seen Obj. Unseen Instr. Row: 57.8 / 43.9 / 48.5. Table 16 total success rate 50.09% (strict 35.19%). Headline variant.

Real Franka Emika Panda; 3 tasks (knocking, covering, pick&place sink); 54 rollouts per model (18 per task) split across seen-obj/unseen-combo, unseen-obj, unseen-instruction. Columns: Seen Obj. Unseen Combo / Unseen Obj. / Seen Obj. Unseen Instr. Row: 43.4 / 31.4 / 35.6.

Real Franka Emika Panda; 3 tasks (knocking, covering, pick&place sink); 54 rollouts per model (18 per task) split across seen-obj/unseen-combo, unseen-obj, unseen-instruction. Columns: Seen Obj. Unseen Combo / Unseen Obj. / Seen Obj. Unseen Instr. Row: 36.5 / 37.4 / 28.1. Pretrained on Something-Something V2 human video only.

Zero-shot (pretrained OpenVLA, no fine-tuning) on WidowX / BridgeData V2 setup; 17 tasks x 10 trials = 170 rollouts; partial success 0.5 allowed on some tasks; baselines in same table: RT-1-X 18.5±2.7%, Octo 20.0±2.6%, RT-2-X 50.6±3.5%

Octo fine-tuned with ~100 target demos per domain (Berkeley Insertion 70, Stanford Coffee 75, CMU Baking 50, Berkeley Pick-Up 60, Berkeley Coke 100, Berkeley Bimanual 80); ~20 trials per domain; baselines ResNet+Transformer scratch 20%, VC-1 15%. Zero-shot results (Fig. 4) are bar charts with no printed values

ALOHA dual-arm; RDT pre-trained then fine-tuned on self-collected ALOHA data; 8 trials per cup (24 total); Total for seen cup 1 = 50, unseen cup 1 = 75; ACT/OpenVLA/Octo Total = 0 for all cups; RDT (scratch) Total = 0

Fine-tuned on real-robot data; Pick 70.8 / Stack 82.3 / Place 60.4; Octo-Base 4.9, OpenVLA 12.1

Multi-task fine-tuning, 50 demos per task, 200k steps; baselines ACT (single-task) 48.3, π0 (multi-task) 61.7

π0 fine-tuned per task by the GO-1 authors on AgiBot G1; 5 tasks, 30 trials per task; bar labels verified from the rendered figure

Run by SmolVLA authors on SO-100 arm

Post-trained by LingBot authors with 130 episodes per task; 100 tasks, 4 platforms (Agibot G1, AgileX, Galaxea R1Pro, Leju KUAVO 4 Pro), 15 trials per task-robot pair; per-platform SR 7.77 / 17.20 / 14.10 / 12.91

AgiBot G1; 5 tasks; 30 trials per task (10 seen setup + 20 with variations/distractors); GO-1 pre-trained on AgiBot World beta then fine-tuned per task (lr 2e-5, batch 768, 30k steps); score = normalized completion averaged over rollouts (1.0 full success, fractional partial). Values are bar labels printed in the figure (no numeric table); attribution verified from the rendered PDF page. Baselines: RDT-1B 0.46, π0 0.58, GO-1 w/o Latent Planner 0.66; Fig. 1 teaser prints 78 / 66 / 46 and '32% gain'

Same protocol; text: latent planner yields 'an average improvement of 0.12 task completion score'

GM-100: 100 tasks, 4 platforms (Agibot G1, AgileX, Galaxea R1Pro, Leju KUAVO 4 Pro), 130 post-training episodes per task, 15 trials per task-robot pair; per-platform SR 11.98 / 18.93 / 20.98 / 15.60; baselines WALL-OSS 3.85, GR00T N1.6 7.31, π0.5 13.00

Per-platform SR 12.82 / 15.50 / 18.89 / 17.59

GM-100 tabletop bimanual benchmark; trial counts not stated in README

Post-trained by LingBot authors with 130 episodes per task; 100 tasks, 4 platforms, 15 trials per task-robot pair; per-platform SR 5.23 / 3.26 / 14.29 / 6.45

Real Fourier GR-1 humanoid; pick one of two fruits and place on plate, target side randomized; trial count not stated

Real Unitree G1 humanoid, 1K post-training demos

Real-world WidowX zero-shot (out-of-the-box, no fine-tuning), 7 BridgeV2-style tasks, 11 trials each (77 rollouts total, Fig. 5). No average printed in the paper; per-task 'Success' column copied.

Generalizability suite: Lightning Variation 66.7, Visual Distractor 53.3, Novel Object 86.7; average score 2.49.

Text statement, not a table; 1000 demonstrations; Pick-Place-Cup (500 demos) reported only as 'above 90%' and OOD Collect-Waste as 'above 80%'. Robot/trial counts not printed in text.

Zero-shot real-world grasping on Franka Panda (no real-world action data); 15 objects x 2 trials per condition per category group, 300 trials total. Conditions: basic / light / background (b.g.) / distractor (dis.) / height. SPL 87.2.

7-DoF Franka Research 3 + 6-DoF Inspire hand, RealSense L515; 20 randomized trials per task; post-training on 50-100 teleop trajectories per task. Caption says 'Success rates (%)' but values are printed as fractions (0-1); copied as printed. Model size not given in the row label. Baselines GR00T N1.5 and InternVL3 (same architecture without physical instruction tuning).

Number stated in text/bar chart, not a table. Averaged across robot platforms (Franka Panda, WidowX, AgiBot G-1, etc.) and task categories; Franka pick-and-place 94.0%.

Real-world 'Sweep Leftover' task ablation; ~100 demos per task, 15 trials per evaluation per protocol text; printed with std. Baselines: From Scratch 43.3±3.8, VC-1 53.3±2.6, No Prop. Finetuned 63.3±2.6. Printed as 76.7±3.3.

Real-world 'Sweep Leftover' task ablation; ~100 demos per task, 15 trials per evaluation per protocol text; printed with std. Baselines: From Scratch 43.3±3.8, VC-1 53.3±2.6, No Prop. Finetuned 63.3±2.6. Printed as 70.0±3.0.

Real robot: Cobot-Magic dual-arm mobile robot with Piper arms (fixed-base bimanual setting). 6 tasks: Stack Bowls 58, Restore Bottle 38, Click Bell 66, Place Vegetable 50, Pack Bottles 28, Lift Pot 25 (%). 100 episodes per task per setting. Baselines: ACT 18.6, TinyVLA 17.5.

Table 2 'Zero-Shot Robot Manipulation': deployed zero-shot on two Franka arms with Robotiq grippers in two different labs; image-goal, closed-loop MPC (CEM) planning with the V-JEPA 2-AC world model; success out of 10 trials per lab with permutations (object location, starting pose); 'Avg' row is the printed average of Lab 1 and Lab 2. Action-conditioned model trained on 62 h of unlabeled DROID data. Octo baseline Avg: Reach 100, Grasp Cup 15, Grasp Box 0, Reach w/ Obj Cup 15, Box 70, Pick-&-Place Cup 15, Box 10.

Metric is a normalized task score (0-1), not binary success; averaged over 10 trials (Fig. 7 protocol). Bimanual AgileX robot. The same 0.92 is obtained with Stage 1+2 only (no task-specific Stage 3). Baselines Octo/OpenVLA/DP scored ~0 (figure-only).

Table 2 'Real Robot Experiment Results'. 7-DoF Kinova Gen2 with wrist RealSense + Kinect Azure. Values printed as fractions. Baselines: RT-1 0.27/0.13/0.00/0.35, MT-R3M 0.15/0.13/0.10/0.30. Number of rollouts per setting not stated in the table.

Franka Research 3 with Robotiq gripper; post-trained on DROID, evaluated zero-shot in a held-out environment with held-out objects; each task evaluated over 10 trials (binary success except one sequential task with partial credit). Average over the DROID zero-shot task set as stated in text.

ATTRIBUTION CAVEAT: Table 2 'Fine-tuning performance of RDT2 and baseline models on challenging real-world tasks' reports the RDT2-UltraFast variant (Sec. 6.3: 'We use the RDT2-UltraFast variant in this experiment'; Appendix: obtained from the fine-tuned RDT2-FM via consistency distillation). It is a sibling of RDT2-VQ, not RDT2-VQ itself; recorded under this id only because the batch id covers the RDT2 family. RDT2-VQ's own zero-shot results (Pick, Pick&Place, Wiping, Shaking, Button Pressing; 256 trials per task) appear only in Figure 3 (bar chart with standard-error bars) and are not printed as numbers. Fine-tuning: from Stage-1 UMI-pretrained checkpoint, frozen VL backbone, 50K steps + 20K distillation steps. Baselines pi0.5 and pi0-FAST fine-tuned on the same data. Trial counts for Table 2 not stated in text. Same row: pi0.5 36, pi0-FAST 29. Sub-tasks: left sleeve 97, right sleeve 95, final fold 81; unseen object 51.

Real-world long-horizon chemistry procedure (MOF) on the authors' bimanual platform; all methods fine-tuned on the same demonstrations; 'multiple trials with randomized object placement' (exact count not stated for MOF; 15 trials per binding for the instruction-following tasks). Per-task numbers for Sort/Insert/Move Tubes are in Figure 8 (figure-only).

Table 1 'Analysis of Pre-training' (Sec. 3.4, GE-Act on AgiBot G1): row VidAW ✓ (init from GE-Base) + VidAda ✓ (task-specific video adaptation), 'w/ S' (robot state as input) column. Text: 'in-domain pretraining achieves 64 SR and 81 E2E, which further improve to 76% and 89% when combined with general video pretraining.' Values printed as fractions. Metrics: E2E = end-to-end success of the complete task (retries of sub-steps allowed); SR = step-wise success (completed sub-steps / total). The table does not name the task(s) or trial counts for this ablation; without-state ('w/o S') values are E2E 0.37 / SR 0.37. GE-Act vs UniVLA / GR00T N1 / pi0 on the five AgiBot G1 tasks, the Agilex Cobot Magic / Dual Franka cross-embodiment tasks and the four RoboTwin tasks are figure-only (Fig. 8, Fig. 11) - no numbers in text. No CALVIN or LIBERO results in the paper (v1-v3) despite the Dec-2025 release of 'GE-Act trained on Calvin' weights.

Cross-domain (driving) adaptation; UniVLA 81.7 in same table

Simulated GR-1 humanoid tabletop tasks

In-house embodied VQA benchmark; baseline Qwen2.5-VL-3B scores 46.1 / 57.7 / 59.8.

Block-spell instruction-following accuracy; WALL-OSS action-only 26 / 80, pi0 action-only 9 / 35. Trial counts not printed.

Task-scheduling success of the VLM planner in real-world long-horizon/reasoning scenarios (not end-to-end manipulation success); compared with Gemini-2.5 Pro, GPT-5, GPT-4o, Qwen2.5-VL.

Comparison with AnyGrasp (91.6); arbitrary grasping common 93.3, transparent 86.6; 5 Hz.

Embodied reasoning benchmarks vs standard VLMs (GPT-4o overall 40.5, Gemini 1.5 Flash 44.1).

Language Table simulation; printed 62.0±8.7 (StdErr). Other columns: In-domain Unseen 49.6±9.5, Cross-task Seen 73.2±6.8 / Unseen 54.8±9.1, Cross-env Seen 33.6±12.7 / Unseen 29.6±12.0.

OmniGibson simulator, R1-Pro embodiment; 50 tasks x 10 instances; 'Task Success Score' is the challenge ranking metric (task progress). Single generalist policy (Table 4 lists 4-epoch checkpoint; 1-epoch checkpoint scores lower).

Zero-shot robustness test: model trained on LIBERO evaluated directly under LIBERO-Plus perturbations. Per-category: Camera 83.1, Robot 55.1, Language 86.9, Light 96.4, Background 98.2, Noise 95.6, Layout 85.2. Baselines: pi0.5 84.4, Cosmos-Policy 82.2.

Official SAIL blog post 'MiniVLA: A Better VLA with a Smaller Footprint' (Belkhale & Sadigh, Dec 12 2024) - the project has no paper; README points here. LIBERO-90 (90-task suite), NOT the LIBERO 4-suite; 'rollout task success rates in simulation averaged across the 90 tasks'; rollout counts not stated. MiniVLA = Prismatic VLM with Qwen2.5-0.5B LLM (+ DINO-SigLIP ViT), ~1B params total; models trained to the stated action-token accuracy. OpenVLA (7B) row in the same tables: 62%. Corroborated by the author's X thread https://x.com/suneel_belkhale/status/1867631572255707529 ('from 77% to 82%'). Base MiniVLA without VQ or multi-image inputs.

Official SAIL blog post 'MiniVLA: A Better VLA with a Smaller Footprint' (Belkhale & Sadigh, Dec 12 2024) - the project has no paper; README points here. LIBERO-90 (90-task suite), NOT the LIBERO 4-suite; 'rollout task success rates in simulation averaged across the 90 tasks'; rollout counts not stated. MiniVLA = Prismatic VLM with Qwen2.5-0.5B LLM (+ DINO-SigLIP ViT), ~1B params total; models trained to the stated action-token accuracy. OpenVLA (7B) row in the same tables: 62%. Corroborated by the author's X thread https://x.com/suneel_belkhale/status/1867631572255707529 ('from 77% to 82%'). VQ action chunking (Residual VQ, horizon 8). Same table: '+ VQ h1 (single action)' 62.4%.

Official SAIL blog post 'MiniVLA: A Better VLA with a Smaller Footprint' (Belkhale & Sadigh, Dec 12 2024) - the project has no paper; README points here. LIBERO-90 (90-task suite), NOT the LIBERO 4-suite; 'rollout task success rates in simulation averaged across the 90 tasks'; rollout counts not stated. MiniVLA = Prismatic VLM with Qwen2.5-0.5B LLM (+ DINO-SigLIP ViT), ~1B params total; models trained to the stated action-token accuracy. OpenVLA (7B) row in the same tables: 62%. Corroborated by the author's X thread https://x.com/suneel_belkhale/status/1867631572255707529 ('from 77% to 82%'). Adds 2-frame image history on top of VQ h8.

Official SAIL blog post 'MiniVLA: A Better VLA with a Smaller Footprint' (Belkhale & Sadigh, Dec 12 2024) - the project has no paper; README points here. LIBERO-90 (90-task suite), NOT the LIBERO 4-suite; 'rollout task success rates in simulation averaged across the 90 tasks'; rollout counts not stated. MiniVLA = Prismatic VLM with Qwen2.5-0.5B LLM (+ DINO-SigLIP ViT), ~1B params total; models trained to the stated action-token accuracy. OpenVLA (7B) row in the same tables: 62%. Corroborated by the author's X thread https://x.com/suneel_belkhale/status/1867631572255707529 ('from 77% to 82%'). Adds wrist-camera image on top of VQ h8. Blog: 'quite close to the state-of-the-art performance on Libero 90, from Baku'.