PaperlayerView exact revision on arXiv

Source-linked AI summary

Rho: A Foundation for Efficiently Adaptable VLA Models

Rho Team, Simran Bagaria, Daphne Chen, Dean Fortier, Jianlong Fu, Michael Harrison, Tess Hellebrekers, Neel Joshi, Andrey Kolobov, Dalton Moore, Galen Mullins, Michael Murray, Eduardo Salinas, Reuben Tan

arXiv:2609.38164v1cs.RO

TL;DR

Existing VLAs and WAMs have low zero-shot success on unseen manipulation tasks and environments, making adaptation a crucial step toward practical use. Rho uses staged training, including embodiment-specific midtraining, to support downstream task adaptation on three bimanual robot platforms. Across the reported evaluations, embodiment midtraining halves the task-finetuning data needed to reach a given success-rate threshold.

  • Problem

    Existing VLAs and WAMs have low zero-shot success on unseen manipulation tasks and environments, while adaptation is a crucial step between pretrained models and practical use.

    Source: §1 · p. 1

  • Method

    Rho combines a physically grounded foundation model with embodiment-specific midtraining and subsequent task adaptation, with an optional latent policy for corrective online updates.

    Source: §1 · p. 2, §6.2 · p. 11

  • Results

    Embodiment midtraining halves the task-finetuning data needed to reach a given success-rate threshold, and Rho is a strong contender across evaluated tasks and embodiments.

    Source: §1 · p. 3, §11 · p. 31

  • Takeaways & Limitations

    Rho’s embodiment-specific models are reusable starting points for adapting to downstream tasks, and its latent policy can continue improving after offline specialization.

    Source: §1 · p. 2

    The evaluations do not establish universal generalization, and the strongest-model claim is limited to the reported baselines, training budgets, and evaluation protocols.

    Source: §9 · pp. 29–30

Abstract

from arXiv · show

General-purpose physical AI models must combine broad visual and linguistic capabilities with precise control across robot embodiments and efficient adaptation to downstream tasks. We introduce Rho, a family of open-weights VLA models for bimanual manipulation designed for data-light task adaptation on 3 embodiments representative of dual-arm robots across research labs and the industry -- YAM Box, UR AI Trainer, and FR3 Duo. We systematically ablate Rho's action-expert architecture and training recipe, and show in controlled simulation and physical-robot experiments that embodiment midtraining improves downstream adaptation. The resulting Rho variants for YAM Box, UR AI Trainer, and FR3 Duo match or outperform existing open-weights VLAs and achieve the strongest overall performance across the tasks, embodiments, and baselines evaluated in this report. We further demonstrate the Rho model family's built-in capacity for online adaptation: an internal latent policy learns from corrective feedback to select observation-conditioned noise inputs for the frozen flow-matching action expert. With as few as 15 corrected episodes, adapting this lightweight module enables Rho to handle task situations at the fringe of its offline finetuning distribution. Together, these results position Rho as both a strong general-purpose robotic manipulation model and a practical foundation for adaptation. We release the base Rho model and the embodiment-specific checkpoints to facilitate Rho's deployment in research experiments and practical industrial use cases.

1 Introduction

Rho targets data-efficient adaptation by separating robot-embodiment learning from task specialization. Its evaluated models combine staged training with embodiment-specific checkpoints and achieve strong performance across the reported settings.

1 Introduction

  • Rho is a 5B-parameter open-weights VLA family for bimanual manipulation, designed for data-efficient adaptation to new environments and tasks.Its premise is that models should master the target robot’s vision, language, and control before task adaptation.Source: §1 · pp. 1–2
  • The training recipe couples a Phi-family VLM backbone to a flow-matching action expert, then progresses from multi-embodiment pretraining through target-embodiment midtraining to task adaptation.The action expert generates continuous action chunks; the midtraining stage teaches target-robot control across tasks, motions, and environments.Source: §1 · p. 2
  • Across the evaluated settings and open-weights baselines, Rho achieves the strongest overall performance.The report examines its action-expert architecture, pretraining mixture and scheduling, and embodiment midtraining.Source: §1 · p. 3
  • The released family includes Rho-base and three embodiment-midtrained variants, intended as reusable starting points for downstream robotic manipulation scenarios.The variants target YAM Box, UR AI Trainer, and FR3 Duo.Source: §1 · pp. 2–3
  • Embodiment midtraining halves the task-finetuning data needed to reach a given success-rate threshold versus finetuning pretrained Rho without midtraining.The report also demonstrates continued improvement from corrective interaction through Rho’s internal latent policy.Source: §1 · p. 3

2 Physically grounded Phi-family backbone

Rho builds its Phi-family vision-language backbone for robot control by pairing a compact visual-language architecture with robotics-relevant physical grounding data.

2.1 Backbone architecture

  • Phi-Phy combines a Phi-family causal decoder, SigLIP 2 NaFlex vision encoder, and two-layer projector that maps visual features into language space for joint visual-linguistic processing.The encoder preserves image aspect ratio; pretraining camera views are padded to 256×256.Source: §2.1 · pp. 3–4

2.2 Physical grounding data

  • The grounding additions target physical perception and reasoning, including public robotics data, repurposed multi-image VQA, and synthetic difference-spotting and odd-one-out tasks.Roughly 6M of 100M SFT records are in these three categories, alongside 30M general caption, perception, object-existence, and UI-heavy grounding data.Source: §2.2 · pp. 4–5

3 Flow-matching action architecture

Rho conditions a continuous flow-matching action expert on visual-language context and robot state, using configurable action representations and a shared backbone–expert interface.

3 Flow-matching action architecture

  • Rho converts grounded visual-language context into continuous action chunks by conditioning a flow-matching expert on observations, instruction, and robot state.The learned field transforms Gaussian noise toward an action chunk; inference uses 10 explicit Euler steps by default.Source: §3 · p. 5, §3.2 · p. 5

3.1 Observation and action representations

  • The interface supports up to 32 state and action dimensions, multiple control parameterizations, and configurable prediction and execution horizons.Proprioception is projected separately from the interleaved image-language sequence; action chunks may be replanned after a shorter prefix or combined with real-time chunking.Source: §3.1 · p. 5

3.2 Action expert

  • The 542-million-parameter expert uses 12 transformer blocks with grouped-query attention and shared timestep modulation to process state and noisy-action tokens against VLM context.Its 2,048-dimensional action representation uses 128-dimensional attention heads.Source: §3.2 · p. 6, Figure 2 · p. 6

3.3 Backbone–expert interface

  • Each action-expert block cross-attends to a projected intermediate Phi-Phy representation, while the VLM context is reused across flow-integration steps.During pretraining, the vision encoder, projector, and decoder blocks remain trainable at a small learning rate, while the action expert and robot-specific projections use a larger one.Source: §3.3 · p. 7

4 Pretraining

Rho pretrains a transferable continuous-control model on a broad, robot-dominated mixture, standardizes heterogeneous robot actions, and retains vision-language supervision during training.

4 Pretraining

  • Multi-embodiment, multitask pretraining aims to form transferable action representations before adaptation to a particular physical platform.The robot data is dominant, with vision-language data included in the mixture.Source: §4 · p. 7, Figure 3 · p. 8

4.1 Pretraining data mixture

  • Rho-Tomyum combines data from more than a dozen embodiments, including public, proprietary physical-robot, UMI-style, and simulated demonstrations, with a smaller vision-language component.The robot data spans diverse tasks, environments, cameras, action spaces, and control frequencies; source weights balance dataset scale and embodiment coverage rather than raw concatenation.Source: §4.1 · pp. 7–8, Figure 3 · p. 8

4.2 Robot data standardization

  • Robot actions use chunk-relative delta end-effector targets with translation, continuous 6D relative orientation, and absolute gripper position, padded to width 32.Targets span about one second at each dataset’s native rate; per-dataset, per-dimension, and per-position percentile normalization clips actions to [−1.5, 1.5].Source: §4.2 · p. 8

4.3 Pretraining procedure

  • Pretraining jointly optimizes the vision encoder, cross-modal projector, language decoder, and action modules while freezing the token-embedding table.Robot and vision-language modules use separate peak learning rates of 1×10−4 and 1×10−6 for newly initialized action modules and pretrained VLM modules, respectively.Source: §4.3 · p. 9

4.4 Vision-language co-training

  • Each update is either a robot flow-matching batch or a vision-language batch, sampled at a 9:1 ratio, with the vision-language cross-entropy loss weighted by 0.02.Vision-language updates bypass robot state and actions; the loss coefficient controls gradient scale separately from update frequency.Source: §4.4 · p. 9

5 Embodiment midtraining

Rho separates learning a target robot’s embodiment from downstream task adaptation, midtraining each platform-specific checkpoint on broad target-platform data.

5.1 Embodiment adaptation

  • Embodiment midtraining adapts pretrained Rho to a target robot’s observations, kinematics, action space, and control conventions across tasks before individual-task adaptation.Each checkpoint trains for two epochs on a broad target-platform multitask dataset while retaining 9:1 robot-to-vision-language co-training; the goal is broad motions, not mastery of one task.Source: §5 · p. 9, §5.1 · p. 10

5.2 YAM Box

  • Rho-YAM-Box uses roughly 300 hours sampled across 65 task families, split approximately evenly between controlled-laboratory and everyday-environment demonstrations.Its task range includes precision assembly and insertion, tool use, deformable objects, and long-horizon bimanual manipulation under varied visual conditions.Source: §5.2 · p. 10

5.3 UR AI Trainer

  • Rho-UR-AI-Trainer midtrains on about 260 hours of physical demonstrations across 28 task families, augmented with roughly 36 hours of successful simulated trajectories across five tasks.The simulated rollouts add structured stacking, placement, and insertion behaviors on the same embodiment and camera setup, with randomized object colors, camera positions, and focal lengths.Source: §5.3 · p. 11

5.4 FR3 Duo

  • Rho-FR3-Duo adapts to the dual-arm Franka platform using about 280 hours of physical demonstrations across 19 task families.The collection spans bimanual and contact-rich behaviors, from broad workspace motions to precise coordination between arms.Source: §5.4 · p. 11

6 Vertical task adaptation

After embodiment midtraining, Rho uses offline task demonstrations or online interaction for vertical specialization. Its online option adapts a lightweight internal latent policy while keeping the vision-language backbone and action generator frozen.

6 Vertical task adaptation

  • Vertical task adaptation follows reusable embodiment midtraining and includes offline learning from task demonstrations and online learning from interaction.Source: §6 · p. 11

6.1 Offline supervised adaptation

  • Offline adaptation fine-tunes pretrained Rho or an embodiment-midtrained checkpoint using only target-task robot demonstrations, without interleaved vision-language batches.Source: §6.1 · p. 11, §6 · p. 11

6.2 Online adaptation in latent action space

  • Online adaptation replaces Gaussian latent sampling with observation-conditioned πϕ(o), updating only this lightweight submodule while the backbone and action generator remain frozen.Source: §6.2 · pp. 11–12
  • Corrective chunks supervise the latent policy by inverting the frozen flow to recover latent targets; FlowDAgger is evaluated in simulation and on FR3 Duo.Source: §6.2 · p. 12

7 Empirical evaluation

Section 7 evaluates Rho across simulation and physical-robot tasks and embodiments, comparing it with existing open-weight VLAs and examining how architecture, pretraining, embodiment midtraining, and latent-space online adaptation affect task adaptation and data efficiency.

7.1 Action architecture ablations

  • A 2,048-wide action expert with 128-dimensional heads reaches 0.673 RoboEval success, while a 12-block compressed expert retains 0.671 ± 0.008 with 542M parameters.Sharing AdaLN modulation maps reduces parameters from 1.649B to 882M while matching baseline performance; grouped-query attention achieves 0.672 with 693M parameters. Layerwise cross-attention scores 0.605 ± 0.008 versus 0.673 ± 0.015 when blocks attend to the same projected hidden sequence.Source: §7.1 · pp. 13–14

7.2 Pretraining recipe ablations

  • At 20k adaptation steps, robot+VL pretraining reaches 0.963 average LIBERO success and 0.623 RoboEval success, versus 0.943 and 0.558 with robot-only pretraining.The selected action-expert pretraining learning rate is 10^-4, reaching 0.673 RoboEval success after 40k adaptation steps versus 0.633 at 2 × 10^-4 and 0.622 at 5 × 10^-4.Source: §7.2 · pp. 14–15, Figure 5 · p. 15, Table 6 · p. 15

7.3.1 Embodiment midtraining and data efficiency of downstream adaptation

  • Embodiment midtraining halves the finetuning data needed for a given success rate across RoboTwin’s tested data budgets until saturation, with a 30% gain over base Rho at 12.5% data.The midtrained policy using 25% of finetuning data performs as well as base Rho using 50%; the effect holds in Easy and Hard evaluation, including scenes with distractor objects.Source: §7.3.1 · p. 16, Figure 6 · p. 16

7.3.2 Midtraining compute and task adaptation efficiency

  • On FR3 Duo plug insertion, success rises from 18/30 trials without midtraining to 23/30 after one epoch and 26/30 after two epochs.All conditions use the same 50k-step task-adaptation budget; reported gains appeared especially on initial plug configurations near the edge of the training distribution.Source: §7.3.2 · p. 17, Figure 7 · p. 17

7.4 Rho and existing VLAs: performance comparison in simulation

  • On RoboEval, Rho leads the task-level success results on 3 of 8 tasks and matches the best baseline on the other five; it also has the highest task progression and shortest reported paths.Across behavioral metrics, π0.5 has the fewest environment collisions, while MolmoAct2 has the lowest bimanual gripper vertical difference, self-collision count, and slip count.Source: §7.4 · pp. 17–18, Figure 8 · p. 18, Table 7 · p. 18
  • 97.9% average LIBERO success and 73.4% RoboEval success make adapted Rho the strongest aggregate model on LIBERO and 6.3 percentage points ahead of the next-best RoboEval baseline.On LIBERO-Long, Rho reaches 95.9%, exceeding the next-best model by 2.7 percentage points; on the three shorter suites it is within 0.5 points of the best reported value.Source: §7.4 · pp. 17–19, Figure 8 · p. 18, Table 8 · p. 19

7.5.1 YAM Box

  • Physical-robot comparisons test embodiment-midtrained Rho variants against π0.5 and GR00T-N1.7 on FR3 Duo, UR AI Trainer, and YAM Box, with MolmoAct2 added for YAM Box.The YAM Box study finetunes models on approximately 2,000 demonstrations across six BusyBox task categories, using parallel A/B evaluation to reduce effects from environment drift and task initial configurations.Source: §7.5 · p. 19, §7.5.1 · pp. 19–20
  • 90% mean success on BusyBox matched π0.5 and exceeded GR00T-N1.7 and MolmoAct2 after task-specific finetuning.The result came from a high-data setting; π0.5 and MolmoAct2 had prior exposure to related dual-arm embodiments, making them strong comparisons.Source: §7.5.1 · pp. 20–21, Figure 10 · p. 21

7.5.2 UR AI Trainer

  • 46.7% success on Toolbox packing and 60.0% on Small-electronics cleanup were the highest observed rates, ahead of π0.5 (20.0%, 40.0%) and GR00T-N1.7 (16.7%, 13.3%).Despite these wins, all models had low absolute success after only about 150–200 demonstrations; on Small-electronics cleanup, Rho and π0.5 had similar mean task progress (82.5% and 80.8%).Source: §7.5.2 · pp. 21–22, Figure 12 · p. 23

7.5.3 FR3 Duo

  • 80.0% mean success and 88.7% mean task progress on FR3 Duo exceeded GR00T-N1.7 (50.0%, 71.3%) and π0.5 (46.7%, 64.5%).Rho-FR3-Duo also had higher observed success and progress than both baselines on every variant, with its largest success margins on Tumbler Stacking (+33.3 points) and test-tube racking (+30.0 points).Source: §7.5.3 · pp. 23–24, Figure 14 · p. 25

7.6 Online-adaptation results

  • 57.5% to 73.3% average success followed latent-space adaptation on six Meta-World tasks using 50 additional episodes per task.Gains included Assembly (45% to 78%) and Stick pull (58% to 78%); Pick Out of Hole improved from 22% to 37%, while Basketball gained 6 points.Source: §7.6 · p. 25, Table 9 · p. 26
  • 9/30 to 21/30 test-tube assembly successes and 20/30 to 28/30 plug-insertion successes followed online adaptation.The FR3 Duo policies began from 150 offline demonstrations per task and adapted using 15 additional human-supervised episodes on 10 hard configurations near the task-training distribution’s edge.Source: Table 10 · p. 26, §7.6 · p. 26

8 Related work

Related work situates Rho among VLA, video-based, and world-action approaches, then distinguishes its grounding, staged data recipe, robot-specific midtraining, and lightweight online adaptation. The paper positions these choices against limited controlled architectural comparisons and prior adaptation strategies.

8.1 Architectures of VLA models

  • VLA action-expert architectures differ in which vision-language features condition action generation; available ablations are limited, and Rho reports little overall benefit from layer-wise cross-attention.Prior comparisons include knowledge insulation [12] [16] [68] and Fang et al. [14]; Rho therefore optimizes the action-expert shape without layer-wise cross-attention.Source: §8.1 · pp. 26–27

8.2 Other architectural paradigms in physical AI

  • VLAs inherit semantic and language grounding from image-text pretraining, whereas video-based policies inherit spatiotemporal dynamics priors but incur higher training costs and inference latency.VAMs attach an action decoder to video-model features, while WAMs jointly predict visual states and actions.Source: §8.2 · p. 27

8.3 Physical grounding and VLA pretraining

  • Rho grounds its Phi-family backbone and retains physical-grounding data during multi-embodiment pretraining and embodiment midtraining; unlike Intelligence et al. [20], it adds grounding data to the VLM training mixture.The paper notes that general VLM knowledge is not necessarily optimized for robot spatial relationships, affordances, action consequences, or observation distributions.Source: §8.3 · pp. 27–28

8.4 Data recipes, normalization and representations

  • Rho uses multi-task, multi-embodiment demonstrations including UMI-style data [10] for pretraining, single-embodiment data for midtraining, and narrower single-embodiment data for finetuning.Its action and proprioception representations use end-effector pose with a continuous 6D rotation representation and chunk-start-relative actions.Source: §8.4 · p. 28

8.5 Embodiment and task adaptation

  • Rho separates broad robot-specific embodiment midtraining from narrower task adaptation, and reports benefits across a wide range of finetuning-data regimes.Unlike π0.5 [20] and MolmoAct2 [14], which use later training to specialize toward zero-shot generalization, Rho prepares for downstream task-specific adaptation.Source: §8.5 · pp. 28–29

8.6 Online robot learning

  • Rho’s internal latent policy replaces random Gaussian sampling with an observation-conditioned initial condition for a fixed action generator, concentrating online updates in a smaller module.The cited prior framing describes online robot learning as a way to improve policies using experience gathered under deployment, including corrective intervention.Source: §8.6 · p. 29

9 Limitations and future work

The evaluation is bounded to reported benchmarks, embodiments, and tasks, and the report identifies language steerability and agentic-system use as open questions. It also points to tactile feedback as a possible future direction.

Evaluation scope

  • The experiments cover a finite set of benchmarks, embodiments, and tasks, so they do not establish universal generalization across hardware, environments, or task distributions.The strongest-model claim is restricted to the reported baselines, training budgets, and evaluation protocols; midtraining benefits may depend on embodiment-data diversity and quality.Source: §9 · pp. 29–30

Language steerability

  • Rho’s ability to follow diverse natural-language instructions and its sensitivity to phrasing were not systematically evaluated.The authors identify linguistic robustness as an open research question.Source: §9 · p. 30

Agentic AI systems

  • The effectiveness of Rho within agentic systems driven by frontier generative models remains an open question in the report.The authors say studying these workflows could inform practical deployment strategy.Source: §9 · p. 30

Tactile feedback

  • Tactile feedback is an underused modality that the authors propose exploring as a potential route to greater manipulation robustness and dexterity.They describe standardization, technical shortcomings, and limited deployment as factors that have constrained tactile-data accumulation.Source: §9 · p. 30

10 Public release and responsible deployment

The release provides pretrained and robot-specific checkpoints, downstream and online adaptation code, and a YAM Box task dataset. The authors also urge supervised online updates with platform-specific safeguards and evaluation gates.

Public release

  • The release includes pretrained Rho and embodiment-midtrained FR3 Duo, UR AI Trainer, and YAM Box checkpoints, plus code for finetuning, online adaptation, and evaluation.It also includes the BusyBox dataset collected on YAM Box.Source: §10 · p. 30
  • Models are available on Hugging Face, and code for downstream finetuning and online adaptation is available on GitHub.The report also provides a Rho project overview.Source: §10 · p. 30

Responsible deployment

  • Online updates should use human supervision, platform-specific safeguards, workspace constraints, and evaluation gates before deployment.The authors note that offline adaptation depends on representative demonstrations and online adaptation adds operational and safety considerations.Source: §10 · pp. 30–31

11 Conclusion

Rho combines a physically grounded backbone with a flow-matching action expert and releases checkpoints for three representative bimanual platforms. The report highlights embodiment midtraining’s contribution to data-efficient adaptation and describes the model and training configurations used.

11 Conclusion

  • Rho is an open-weights 5B-parameter VLA family for dual-arm manipulation, with released embodiment-midtrained checkpoints for YAM Box, FR3 Duo, and UR AI Trainer.Across the reported experiments, Rho is presented as a strong contender to state-of-the-art VLAs across a broad range of tasks and embodiments.Source: §11 · p. 31
  • Embodiment midtraining halves the offline finetuning data needed for a given success rate, or routinely improves success by 20% or more after finetuning on a small dataset.The conclusion also reports continued online adaptation of task-specialized policies in latent action space.Source: §11 · p. 31

A Model architecture details

  • The full pretraining configuration couples the Phi-Phy backbone to a 12-block continuous action expert conditioned on decoder-block-14 representations; prediction and execution horizons are configurable.The widths, attention layout, and conditioning interface remain fixed.Source: A Model architecture details · p. 41

B.1 Multi-embodiment pretraining

  • Multi-embodiment, multi-task pretraining combines diverse environments and data sources to form a general continuous-control VLA before adaptation to a particular platform.Table 16 is identified as the multi-embodiment pretraining recipe.Source: Table 16 · p. 43

B.2 Embodiment midtraining

  • FR3 Duo embodiment midtraining uses robot and vision-language batches in a 9:1 ratio and optimizes the complete model to learn the platform’s visual and control conventions.The stage continues from the multi-embodiment checkpoint using a broad platform-specific robot dataset while retaining a small fraction of vision-language supervision.Source: §B.2 · p. 43, Table 17 · p. 44

B.3 Task finetuning

Task finetuning begins from embodiment-midtrained Rho checkpoints and uses demonstrations for the target task family. Table 18 gives a representative FR3 Duo recipe.

B.3 Task finetuning

  • Finetuning starts from the embodiment-midtrained checkpoint and uses only demonstrations for the target task family.Each model is adapted independently on the corresponding task demonstrations.Source: §B.3 · p. 44
  • Evaluation-specific data budgets are reported in Section 7.5.Source: §B.3 · p. 44
  • Table 18 summarizes the FR3 Duo test-tube configuration as a representative task-finetuning recipe.Source: Table 18 · p. 44, §B.3 · p. 44

C Simulation evaluation protocols

The simulation protocols evaluate task adaptation across LIBERO, RoboEval, and RoboTwin, with specified task sets, budgets, and success measures. RoboTwin also tests midtraining on held-out tasks under Easy and Hard conditions.

C Simulation evaluation protocols

  • Success is binary, and task scores are averaged uniformly so longer-horizon or demonstration-rich tasks receive no extra weight.A rollout ends when the benchmark predicate is met or the time limit is reached.Source: C Simulation evaluation protocols · p. 44
  • LIBERO evaluation covers all 40 tasks in four suites, with suite scores averaging ten task success rates and the overall score averaging the four suites.The reported mean spans seeds 42–44 and 6,000 trials; task horizons vary by suite.Source: C Simulation evaluation protocols · p. 45, Table 20 · p. 45
  • RoboEval jointly adapts across eight bimanual tasks and reports success over three evaluation passes, totaling 2,400 trials per model.Each rollout has a 250-step limit; task progression and behavioral measures do not alter the binary success definition.Source: C Simulation evaluation protocols · pp. 45–46
  • The RoboTwin 2.0 study midtrains on 45 UR5 tasks and holds out five tasks for finetuning and evaluation.The 5-task finetuning set has 2,750 demonstrations; midtraining uses roughly 23,000 demonstrations from 45 tasks.Source: C Simulation evaluation protocols · p. 46
  • RoboTwin 2.0 reports Easy and Hard evaluation modes: Easy keeps finetuning backgrounds, lighting, and table height without extra clutter, while Hard randomizes backgrounds and lighting, adds clutter, and varies table height by up to 3 cm.Hard results average over base seeds 10000, 11000, and 13000; Easy results use seeds 42–45. Dataset-size comparisons use 12.5%, 25%, 50%, and 100% finetuning subsets and evaluate each model with 50 trials per task per seed, averaging over three seeds.Source: C Simulation evaluation protocols · p. 46

D Physical-robot evaluation protocols

Physical-robot evaluations use matched initial states and fixed operating conditions to compare Rho across three bimanual platforms. The section also describes how corrective feedback trains an internal latent policy while preserving the pretrained backbone and action generator.

D Physical-robot evaluation protocols

  • Parallel A/B evaluation compares models from shared initial states, with blocks run in sequence to reduce variance and minimize environment drift.Object positions, orientations, and sometimes identities or arrangements vary across blocks, while cameras, controllers, instructions, and workspace stay fixed within each comparison.Source: D Physical-robot evaluation protocols · pp. 46–47
  • UR AI Trainer and FR3 Duo use 30 evaluation blocks per multi-stage task, while YAM Box pools BusyBox demonstrations and uses 60 blocks across task classes.YAM Box overall scores weight task variants equally.Source: D Physical-robot evaluation protocols · p. 47

E.1 YAM Box

  • YAM Box uses two six-degree-of-freedom arms, three RGB views, and predicts bimanual Cartesian targets at 30 Hz.Its adjustable arm spacing creates a shared central reach region for BusyBox manipulation.Source: §E.1 · p. 47

E.2 UR AI Trainer

  • UR AI Trainer uses two UR5e arms, three RGB streams, and an 18-Hz control loop to command Cartesian poses in a shared workspace.The midtraining data also includes successful simulated OmniReset trajectories for stacking, placement, and insertion tasks.Source: §E.2 · p. 48

E.3 FR3 Duo

  • FR3 Duo uses two seven-degree-of-freedom arms and predicts bimanual Cartesian targets at 30 Hz, with pose targets constrained to reachable configurations.The shared workspace supports bimanual handoffs, assembly, and object manipulation.Source: §E.3 · p. 48

F Online-adaptation details

  • Online adaptation trains only Rho’s internal latent policy from corrective interaction, leaving the vision-language backbone and flow-matching action expert frozen.Executed action chunks are inverted through the frozen flow to produce latent regression targets; non-injective flow inversion recovers a reproducing preimage, which is discarded if round-trip MSE exceeds 10^-3.Source: F Online-adaptation details · pp. 48–49

G RoboEval behavioral metrics

The RoboEval behavioral-metrics section provides task-level outcome and behavior breakdowns after 40k adaptation steps. The authors’ contributions span the training recipe, software and evaluation infrastructure, and robot-specific data and experiments.

G RoboEval behavioral metrics

  • Table 22 provides task-level RoboEval outcome and behavioral metrics after 40k task-adaptation steps, making task-specific differences visible.Cells pool three 100-episode evaluations, path lengths sum both arms, and arrows mark preferred directions.Source: G RoboEval behavioral metrics · p. 50, Table 22 · p. 51

H Authors’ contributions

  • Additional contributions included integrating and running online-adaptation experiments, maintaining robot hardware, collecting data, and developing the Phi-Phy vision-language backbone.Source: H Authors’ contributions · pp. 51–52
  • Mullins built core software, training, and evaluation infrastructure and led UR AI Trainer physical experiments; Bagaria contributed simulation data generation and robot demonstrations.Source: H Authors’ contributions · pp. 51–52
  • Murray led Rho’s training-recipe and architecture design, while Kolobov led the project, its midtraining strategy, and evaluation design.Source: H Authors’ contributions · p. 52
Loading 2609.38164v1…