ManiFlow: A General Robot Manipulation Policy via Consistency Flow Training

• ManiFlow is presented as a flow matching model for complex manipulation, including bimanual dexterous tasks.Source: Figure 1 · p. 1
• Panels a–d show autonomous pouring and grasping diverse objects before placing them into containers.Source: Figure 1 · p. 1
• Panel e shows a bottle being passed from one hand to the other.Source: Figure 1 · p. 1

• Existing flow-matching policies remain limited in efficiency, robustness, and generalizability on complex real-world dexterous tasks.Challenges include multi-fingered interactions, temporal coherence, unseen scenarios, and modeling visual, language, and proprioceptive inputs.Source: §1 · pp. 1–2
• Consistency training encourages a consistent mapping from noisy samples to the target distribution, straightening flow paths for accurate actions with fewer inference steps and no pretrained teacher.The paper also reports that beta and continuous-time sampling benefit flow matching and consistency training.Source: §1 · p. 2
• DiT-X uses cross-attention for high-dimensional visual and language inputs and AdaLN-Zero conditioning for low-dimensional inputs, selectively modulating cross-attention features.Source: §1 · p. 2, Figure 2 · p. 3
• ManiFlow consistently improves over diffusion and flow-matching policies in image-based 2D and point-cloud 3D settings, by 45.6% and 11.0%, respectively, and achieves 31.4% improvement in multi-task settings.Separately, ManiFlow improves over π0 by 58% on 4 robustness-test tasks and more than doubles 3D Diffusion Policy success across 8 real-world tasks.Source: §1 · p. 2
• ManiFlow is evaluated on single-arm, bimanual, and humanoid robot setups, with reported robustness to novel objects and environmental variations.Source: §1 · p. 2

Continuous-time Consistency Training

• Consistency training maps intermediate trajectory points toward the target data point and constrains predictions at sampled points to agree.The model takes a step-size argument, and an EMA flow model helps estimate the target velocity.Source: Continuous-time Consistency Training · p. 3, Figure 3 · p. 4

2.1 ManiFlow Training

• ManiFlow combines flow-matching and continuous-time consistency losses, using a zero step-size argument for local instant-velocity estimation.Continuous step-size sampling avoids the bias associated with discrete-time objectives, while the EMA model stabilizes training.Source: Continuous-time Consistency Training · pp. 3–4, §2.1 · p. 3

Time Space Sampling Strategy

• Beta timestep sampling, which emphasizes noisier actions, performs particularly well for robotic control; continuous step sizes also improve performance.The study compares uniform, logit-normal, mode, CosMap, and beta timestep schedules.Source: Time Space Sampling Strategy · p. 4

2.2 Perception

• The 3D encoder preserves point-wise features rather than pooling them, retaining geometric detail for precise object interaction and spatial reasoning.The preferred point density varies by scene: 128 points for cropped, well-calibrated scenes and 4096 for uncalibrated egocentric views.Source: §2.2 · p. 4

2.3 ManiFlow Policy Architecture

• DiT-X uses adaptive cross-attention for visual and language tokens, modulated by AdaLN-Zero scale and shift parameters from low-dimensional inputs.The modulation also adjusts cross-attention inputs and outputs; the design adds modest computational overhead.Source: §2.3 · p. 5, Figure 4 · p. 5
• 31.4% and 34.9% relative improvements: ManiFlow beats 3D diffusion and flow-matching policies across difficulty levels on 48 MetaWorld tasks.Table 1 also reports superior performance on image and point-cloud inputs across 12 dexterous tasks in three benchmarks.Source: Figure 6 · p. 6, Table 1 · p. 6
• DiT-X converges faster and performs better than the version without cross-attention AdaLN-Zero conditioning.The comparison covers training action error and success rate on 10 language-conditioned MetaWorld tasks.Source: Figure 5 · p. 5, §2.3 · p. 6

3.1 Simulation Experiments; 3.3 Real World Experiments

• The evaluation covers 12 dexterous benchmark tasks, 48 language-conditioned MetaWorld tasks, four RoboTwin 2.0 tasks, and eight real-robot tasks across three setups.Baselines vary by experiment and include 2D image and 3D point-cloud policies, as well as π0 for RoboTwin 2.0 robustness and scaling tests.Source: §3.1 · p. 6, §3.3 · p. 9

3.2 Key Findings; 3.3 Real World Experiments

• ManiFlow beats diffusion and flow baselines on three dexterous benchmarks (2D gains: 43.4%, 45.6%; 3D: 15.9%, 11.0%) and averages 98.3% relative improvement over DP3 on real robots.The real-robot comparison covers eight tasks across three platforms.Source: §3.2 · p. 6, §3.3 · p. 9, Table 2 · p. 10
• Robustness tests include novel objects, varied backgrounds, clutter and distractors, lighting changes, and altered table heights; real-world tests also cover unseen object types and geometries.The real-world passage reports no significant performance drop under these changes.Source: Figure 8 · p. 7, §3.2 · p. 8, §3.3 · p. 9

3.2 Key Findings

• On 48 MetaWorld tasks, ManiFlow improves over 3D diffusion and flow baselines by 31.4% and 34.9% on average, and by 125% and 73.6% on very hard tasks; DiT-X learns faster and performs better.The DiT-X comparison covers ten language-conditioned MetaWorld tasks.Source: §3.2 · pp. 7–8
• From scratch with 50 domain-randomized demonstrations, ManiFlow achieves a 58% average relative gain over π0 across four RoboTwin 2.0 bimanual tasks.The evaluation includes novel objects and backgrounds; the reported task-level comparisons are 64.7% vs 24.3% on Lift Pot and 55.5% vs 18.0% on Pick Dual Bottles.Source: §3.2 · pp. 7–8, Figure 7 · p. 7
• On Lift Pot, ManiFlow rises from about 10% success with 10 demonstrations to 97.7% with 200 and 99.7% with 500; π0 reaches 94.0% with 500.The comparison indicates stronger performance in the low-data regime and as demonstration counts increase.Source: Figure 7 · p. 7, §3.2 · pp. 8–9
• On five RoboTwin bimanual dexterous tasks, ManiFlow reaches 63.7% and 64.5% success with one and two inference steps, versus 42.7% and 48.1% for baselines using ten steps.The baselines are 3D Diffusion and Flow Matching Policies.Source: §3.2 · p. 9

3.3 Real World Experiments

• On the real-robot handover task, ManiFlow succeeds in 22 of 30 runs (73%), compared with DP3’s 14 of 30 runs (47%); humanoid pouring improves from 20% to 65%.The tasks involve coordination between dexterous hands and precise bottle control.Source: §3.3 · p. 9

Generative Models for Policy Learning

• Diffusion Policies model multimodal action distributions, Consistency Policies use teacher-student distillation for faster inference, and flow matching has shown improved policy performance and training efficiency [1] [7] [6] [3].The paper identifies limitations of these methods in modeling complex, high-dimensional dexterous behaviors.Source: §4 · p. 10

Visual Imitation Learning

• Many 3D visual imitation methods remain restricted to 6-DoF end-effector control or coarse keypoint prediction, limiting their suitability for highly dynamic dexterous tasks [26] [27] [28] [29] [30] [31] [32].3D Diffusion Policy is cited as a dexterous-task method, while ManiFlow aims to learn robust skills from either 2D or 3D observations.Source: §4 · p. 10

Architecture for Multi-modality Conditioning

• Prior DiT and MDT approaches use transformer or cross-attention conditioning, while ManiFlow adds AdaLN-Zero conditioning to cross-attention with learned scale and shift parameters [33] [34] [9].This modification is intended to make multimodal feature conditioning more flexible and efficient.Source: §4 · p. 10

• ManiFlow combines a continuous-time consistency objective, a time-sampling strategy, and DiT-X, and achieves a 98.3% relative improvement over existing approaches in challenging real-world bimanual dexterous manipulation.The evaluation spans 64 simulation tasks and eight real-world scenarios.Source: §5 · p. 11

• Real-world task success depends heavily on the quality and diversity of training demonstrations.The authors suggest reinforcement learning as a possible way to reduce the burden on demonstration data.Source: §6 · p. 11
• The design choices are inspired by dexterous manipulation, but the authors believe they could also benefit navigation or mobile manipulation.Source: §6 · p. 11
• The paper has only begun exploring multimodal extensions such as tactile input and VLM-based conditioning through points, trajectories, or bounding boxes.Source: §6 · p. 11

A.1 Perception

• 32.9% relative improvement on 2D image inputs and 16.2% on 3D point cloud inputs are reported across dexterous manipulation benchmarks.For 3D perception, the lightweight PointNet-based encoder preserves point-wise features; point-cloud density varies from 128–256 points in calibrated, cropped scenes to 2048–4096 in cluttered scenes.Source: §A.1 · pp. 14–15
• ManiFlow uses ResNet-18 for RGB images, a point-wise PointNet-based encoder for 3D, frozen T5 embeddings for language, and a two-layer MLP for proprioception.The image encoder is trained from scratch; proprioceptive masking is used to reduce reliance on state inputs and shortcuts around visual understanding.Source: §A.1 · pp. 14–15, §A.2 · p. 15

A.2 ManiFlow & Baseline Model Details.

• Action horizons are 4 steps for short-horizon simulation, 16 for bimanual RoboTwin, and 64 for real-world tasks, where temporal ensembling smooths predictions.All tasks use a two-timestep observation history; ManiFlow uses U-Net diffusion/flow-matching baselines.Source: §A.2 · pp. 15–16

A.3 ManiFlow Training Details.

• Training jointly optimizes flow matching and consistency without a pretrained teacher, using a 75%/25% batch split and different timestep sampling strategies.Flow matching uses Δt = 0 and Beta sampling; consistency training samples Δt continuously and uses discrete uniform timestep sampling. Relative Δt conditioning performs better than absolute target-time conditioning, and an EMA model supplies stable consistency targets.Source: §A.3 · p. 16, training · p. 16

A.4 Failure Cases.

• ManiFlow fails on tasks needing detailed contact information and precise force feedback, including delicate assembly and compliant insertion.The authors attribute this boundary to a kinematic-control focus and the absence of tactile sensing and force control.Source: §A.4 · p. 16

B.1 Training Details.

• Models are compared under identical conditions, with mean success rates and standard deviations reported across three training seeds.RoboTwin uses 2,000 epochs and 100 evaluation episodes; Adroit and DexArt use 3,000 epochs, evaluations every 50 epochs over 20 episodes, and the average of the top five success rates.Source: §B.1 · p. 17
• MetaWorld evaluation focuses on the more challenging language-conditioned multi-task setting because both the baseline 3D diffusion policy and our method achieve approximately 90%–100% success in single-task settings for most tasks, limiting meaningful differentiation.Training uses 10 demonstrations per task for Adroit and MetaWorld, 50 for RoboTwin, and 100 for DexArt.Source: §B.1 · pp. 16–17

B.2 Simulation Benchmark

• The simulation suite combines MetaWorld single-arm tasks, Adroit dexterous-hand tasks, DexArt unseen articulated-object tasks, and RoboTwin bimanual tasks.The benchmark selection spans different environments, task complexities, and skill sets.Source: §B.2 · p. 17

B.3 Ablation

• 79.0% success with 500 demonstrations using point-cloud coordinates only exceeds the 3D diffusion baseline’s 32.7%; adding RGB raises ManiFlow to 92.3%.On Pick Apple Messy, ManiFlow rises from 3.7% with 10 demonstrations to 57.3% with 100; with RGB it reaches 86.0% at 100 demonstrations.Source: §B.3 · pp. 17–18
• The time-scheduler ablation favors Beta sampling for timestep t and continuous over discrete sampling for Δt across seven Adroit and MetaWorld tasks.The authors link Beta sampling’s performance to its emphasis on lower, higher-noise timesteps.Source: §B.3 · p. 18, Table 3 · p. 18
• ManiFlow outperforms Diffusion Policy, Rectified Flow, Consistency-FM, and the shortcut model across diverse manipulation scenarios.The comparison uses the same encoder and DiT-X architecture across variants.Source: §B.3 · p. 18, Table 5 · p. 19
• Single-step ManiFlow in 3D-DA scores 3.67 average sequence versus 3.35 for the original 25-step DDPM, with 25× inference speedup.For four-instruction chains, success is 73.0% versus 53.3%; for five-instruction chains, 65.7% versus 41.2%.Source: §B.3 · p. 18, Table 6 · p. 19

C.1 Real-World Setups

• The real-world evaluation uses a Franka single-arm setup, bimanual xArms, and a Unitree H1 humanoid, and covers eight manipulation tasks.The H1 setup adds active head perception and anthropomorphic hands; the setups differ in viewpoint and hardware complexity.Source: §C.1 · pp. 19–21

C.2 Task Descriptions

• The eight real-world tasks cover humanoid grasp-and-place and pouring, bimanual handover, pouring, toy grasping and sorting, and single-arm cap hanging and pouring.Bimanual Toy Grasping is a single-arm task performed within the bimanual setup.Source: §C.2 · p. 21

C.3 Data Collection

• Bimanual demonstrations use Apple Vision Pro teleoperation with approximately 50 demonstrations per task; single-arm tasks use Oculus VR with 70–80 per task.Single-arm data collection varies object type, location, and orientation.Source: §C.3 · pp. 21, 23

C.4 Evaluation Metrics

• Generalization is assessed on seen objects, unseen object types, and distractor perturbations, with most tasks evaluated over 305 rollouts.The exception is Bimanual Sorting; the 305-rollout total covers 29 objects at 10 trials each in humanoid and bimanual settings, plus 15 single-arm trials.Source: §C.4 · p. 23, Table 9 · p. 23

C.5 Evaluation Details for Bimanual and Humanoid tasks

• Bimanual object sets vary bottle geometry and size for pouring and handover, toy shape for distractor-heavy grasping, and cube and cylinder geometry for sorting.The task descriptions specify pouring into a target bottle, handing a bottle to the right hand for placement, and placing a toy into a basket.Source: §C.5 · pp. 23–24, Figure 17 · p. 24
• Bimanual tests use defined workspace regions, including 10.5 cm × 10.5 cm for the target bottle and 15 cm × 15 cm for the pouring bottle.The handover bottle region is 15 cm × 15 cm, with box displacement perturbations of approximately 1.5 cm; toy grasping uses a 21 cm × 21 cm region and basket displacement of around 2.5 cm.Source: Figure 18 · p. 24
• Humanoid tests vary bottle and cup placement directions for pouring and vary object position around the robot for grasp-and-place.The selected humanoid objects also vary in shape and scale.Source: §C.5 · p. 24, Figure 20 · p. 25, Figure 19 · p. 25
• Single-arm pouring randomizes bottle locations within 10 cm × 20 cm and cup locations within 8.5 cm × 20 cm; cap hanging varies cap orientation by 10–20 degrees.The setup uses three bottle types for pouring and two cap types for cap hanging.Source: §C.5 · p. 26

D ManiFlow Training Algorithms

• The appendix lists beta, logit-normal, mode, and cosmap timestep samplers; beta sampling uses α = 1.0, β = 1.5, and s = 0.999.The logit-normal sampler specifies m = 0.0 and s = 1.0, while mode sampling specifies s = 1.29.Source: D ManiFlow Training Algorithms · p. 26
• Flow-matching training interpolates between noise and data and uses their difference, x1 − x0, as the direct target velocity.The time-step increment is set to zero for this training branch.Source: D ManiFlow Training Algorithms · p. 26
• Consistency training estimates a target velocity by integrating an EMA-model velocity from a later interpolated point, then averaging the displacement from the current point.The algorithm samples time intervals and uses the resulting target in the velocity-prediction loss.Source: D ManiFlow Training Algorithms · p. 26
• The training loop updates model parameters by minimizing squared error between the predicted and target velocities.Source: D ManiFlow Training Algorithms · p. 26