PaperlayerView exact revision on arXiv

Source-linked AI summary

RLDX-1 Technical Report

Dongyoung Kim, Huiwon Jang, Myungkyu Koo, Suhyeok Jang, Taeyoung Kim, Beomjun Kim, Byungjun Yoon, Changsung Jang, Daewon Choi, Dongsu Han, Donguk Lee, Heeseung Kwon, Hojin Jeon, Jaehyun Kang, Jaekyoung Bae, Jihyuk Lee, Jimin Lee, John Won, Joonwoo Ahn, Junhyeong Park, Junyoung Sung, Kyungmin Lee, Minseong Han, Minsung Yoon, Sejune Joo, Seonil Son, Seungcheol Park, Seunggeun Cho, Seungjun Moon, Seungku Kim, Yonghoon Dong, Yongjin Cho, Youngchan Kim, Chang Hwan Kim, Dohyeon Kim, Heecheol Kim, Heewon Lee, Hensen Ahn, Hyungkyu Ryu, Hyunsoo Choi, Hyunsoo Shin, Jaeheon Jung, Jaewoo Kim, Jinwook Kim, Joochul Chang, Joonsoo Kim, Junghun Park, Jungwoo Park, Junho Cho, Junhyeok Park, Junwon Lee, Kangwook Lee, Kwanghoon Kim, Kyoungwhan Choe, Manoj Bhadu, Nayoung Oh, Sangjun Kim, Sangwoo Kim, Seunghoon Shim, Seunghyun Kim, Seungjun Lee, Seungyup Ka, Sungryol Yang, Wook Jung, Yashu Shukla, Yeonjae Lee, Yeonwoo Bae, Jinwoo Shin

arXiv:2605.03269v2cs.ROcs.AIcs.LG

TL;DR

Existing VLAs remain limited on complex real-world manipulation requiring motion awareness, long-term memory, and physical sensing. RLDX-1 integrates these capabilities through a unified architecture and outperforms frontier VLAs, achieving 86.8% success on ALLEX humanoid tasks versus around 40% for π0.5 and GR00T N1.6.

  • Problem

    Existing VLAs emphasize versatile intelligence but remain limited on complex real-world manipulation requiring motion awareness, long-term memory, and physical sensing.

  • Method

    RLDX-1 combines MSAT’s modality-specific streams with joint self-attention, synthetic rare-scenario data, three-stage training, and inference optimization.

  • Results

    RLDX-1 consistently outperforms baselines across simulation and real-world tasks, achieving 91.7% success on ALLEX Object-in-Box Selection.

  • Takeaways & Limitations

    The results support RLDX-1 as a general-purpose VLA for human-like dexterous manipulation in real-world environments.

Abstract

from arXiv · show

While Vision-Language-Action models (VLAs) have shown remarkable progress toward human-like generalist robotic policies through the versatile intelligence (i.e. broad scene understanding and language-conditioned generalization) inherited from pre-trained Vision-Language Models, they still struggle with complex real-world tasks requiring broader functional capabilities (e.g. motion awareness, long-term memory, and physical sensing). To address this, we introduce RLDX-1, a general-purpose robotic policy for dexterous manipulation built on the Multi-Stream Action Transformer (MSAT), an architecture that unifies these capabilities by integrating heterogeneous modalities through modality-specific streams with cross-modal joint self-attention. RLDX-1 further combines this architecture with system-level design choices, including data synthesis for rare manipulation scenarios, learning procedures specialized for human-like manipulation, and inference optimizations for real-time deployment. Through empirical evaluation, we show that RLDX-1 consistently outperforms recent frontier VLAs (e.g. $π_{0.5}$ and GR00T N1.6) across both simulation benchmarks and real-world tasks that require broad functional capabilities beyond general versatility. In particular, RLDX-1 shows superiority in ALLEX humanoid tasks by achieving success rates of 86.8% while $π_{0.5}$ and GR00T N1.6 achieve around 40%, highlighting the ability of RLDX-1 to control a high-DoF humanoid robot under diverse functional demands. Together, these results position RLDX-1 as a promising step toward reliable VLAs for complex, contact-rich, and dynamic real-world dexterous manipulation.

1. Introduction

RLDX-1 targets dexterous manipulation beyond versatile visual-language intelligence by integrating motion awareness, long-term memory, and physical sensing into a unified robotic policy. Its system combines the MSAT architecture with synthetic-data augmentation, staged training, and inference optimization, and it outperforms major baselines on reported versatile-intelligence tasks.

  • System Design: RLDX-1 combines a unified multimodal architecture, synthetic data for rare scenarios, three-stage training, and inference optimization for real-time control.The inference pipeline uses static graph conversion and operator fusion, while synthetic data is filtered for motion consistency.
  • Neural Architecture: RLDX-1 addresses motion awareness, long-term memory, and physical sensing as functional capabilities missing from standard versatile intelligence.Its modules process video dynamics, maintain queues of historical observation features, and ingest tactile or torque signals while predicting future physical sensory signals.
  • Neural Architecture: The Multi-Stream Action Transformer assigns each modality a dedicated stream and couples streams through joint self-attention for action generation.This design preserves modality-specific representations while allowing multimodal information to contribute jointly to the predicted actions.
  • Evaluation & Analysis: The evaluation spans simulation benchmarks and real-world humanoid and single-arm manipulation tasks, comparing RLDX-1 with recent VLAs including GR00T N1.6 and π0.5.The benchmarks assess broad VLA capabilities, while real-world tasks assess versatile intelligence and additional functional capabilities.
  • Evaluation & Analysis: RLDX-1 outperforms π0.5 on Unseen Object (37.5% to 54.2%) and Unseen Task (45.8% to 54.2%) in versatile-intelligence tasks.The introduction reports that the performance gap becomes even more pronounced on functional-capability evaluations involving motion awareness, long-term memory, and physical sensing.

2. Neural Architecture

RLDX-1 combines a temporally aware VLM, explicit long-term memory, and an MSAT-based action model to process heterogeneous inputs for dexterous manipulation. Its architecture adds cognition tokens, motion modeling, and dedicated modality streams with cross-modal joint self-attention.

  • 2.1 Temporally Aware VLM: RLDX-1-VLM encodes video and language with Qwen3-VL 8B, adapting it through robot-related VQA training, cognition tokens, and intermediate-layer features for action decoding.The cognition tokens are learnable queries appended to video and language tokens to extract action-relevant representations.
  • 2.1 Temporally Aware VLM: The VLM models motion by combining multi-frame observations with a temporal module that computes space-time self-similarity and updates visual features residually.The module is inserted after the 9th of 27 vision-encoder layers, while the language backbone temporally aggregates and compresses past observations.
  • 2.1 Temporally Aware VLM: An explicit memory module caches recent cognition features at action-chunk intervals, providing historical context when short-term observations cannot reveal task state or progress.The memory queue stores the last nmem cached cognition tokens after the VLM.
  • 2.2 Multimodal Action Model: The action model generates H + 1 future actions using a flow-matching Diffusion Transformer conditioned on cognition, memory, proprioception, and available physical signals.It trains a neural vector field to move noisy action chunks toward clean actions and uses Euler integration for denoising.
  • 2.2 Multimodal Action Model: MSAT processes cognition, action, and optional physical-signal inputs through dedicated streams while enabling cross-modal interaction via joint self-attention.Early blocks use separate cognition and action streams, optionally adding a physics stream; later blocks merge streams for joint processing.

3. Training Data

RLDX-1 is trained on complementary real-world and synthetic robot data spanning single-arm, dual-arm, and humanoid embodiments. The pipeline combines broad public datasets with tactile- and torque-augmented in-house collection and synthetic augmentation for specialized manipulation scenarios.

  • 3.1 Real-World Data: RLDX-1 curates public manipulation datasets across single-arm grippers, dual-arm robots, and humanoids to establish an action prior across diverse embodiments.The collection includes OXE, DROID, Galaxea Open-World, Agibot World, Fourier ActionNet, and Humanoid Everyday.
  • 3.1 Real-World Data: Public data spans over 1M OXE trajectories across 20 embodiments, 92K DROID trajectories across 564 scenes and 86 tasks, and 100K Galaxea trajectories across 150 task categories.Agibot World contributes over 1M trajectories across 217 tasks, while Fourier ActionNet contributes 30K bimanual trajectories for humanoids with dexterous hands.
  • 3.2 In-House Data: In-house data expands physical sensing through an FR3 platform with tactile and torque sensors and an ALLEX humanoid with torque feedback for high-DoF dexterous manipulation.The FR3 setup uses an AnySkin gripper sensor, joint torque measurements, and VR end-effector teleoperation; ALLEX has 7-DoF arms, 15-DoF hands, a 2-DoF waist, a 2-DoF neck, stereo cameras, and estimated joint torques.
  • 3.3 Synthetic Data: The synthetic framework generates action labels with an inverse dynamics model and retains samples whose simulated rollouts are consistent with the generated videos.Data generation combines scene and task augmentation, while filtering uses IDM replay in simulation and a motion-consistency classifier.
  • 3.3 Synthetic Data: Synthetic augmentation targets specialized scenarios that are difficult to scale through direct collection, applying video generative models to public GR-1 and in-house ALLEX datasets.The framework diversifies task instructions with VLMs and scene visuals through image-to-image and video-to-video transformations, then filters generated samples for video quality and motion consistency.

4. Training Procedure

RLDX-1 is trained through pre-training, embodiment-specific mid-training, and task-specific post-training. The procedure expands functionality beyond general manipulation through synthetic and teleoperated data, adaptive demonstrations, and reinforcement-learning refinement.

  • Overview: RLDX-1 uses three stages: multi-embodiment pre-training, embodiment-specific mid-training, and task-specific post-training for refined performance.Pre-training learns general action prediction, mid-training adds embodiment expertise and capabilities, and post-training specializes the policy for tasks.
  • Pre-training: 150K synthetic GR-1 humanoid episodes supplement heterogeneous pre-training data spanning single-arm, dual-arm, and dexterous-hand humanoid embodiments.The mixture is intended to encourage embodiment-agnostic representations while addressing humanoid data scarcity.
  • Mid-training: Mid-training specializes RLDX-1 on ALLEX and FR3 while adding motion awareness, long-term memory, and physical sensing.ALLEX is a 48-DoF humanoid, while FR3 is a single-arm gripper robot.
  • Mid-training: ALLEX training mixes in-house teleoperation with 72K synthetic episodes at 5:5, while FR3 combines 92K DROID episodes with in-house data at 8:2.In-house data exclusively supervises newly added memory, torque, and, for FR3, tactile inputs.
  • Post-training: Adaptive data collection alternates base demonstrations with failure-focused refinement to improve policies on target real-world tasks.The refinement stage expands scenario definitions or variance factors and collects demonstrations targeting observed weaknesses.
  • Post-training: A text-based VLM critic predicts values autoregressively with native number tokens, supporting RECAP-based reinforcement-learning refinement without jointly optimizing critic and policy.RECAP alternates policy rollouts, dataset expansion, and improvement of both components; the text critic addresses adaptation challenges of newly initialized value heads [Tan et al., 2025; Amin et al., 2025; Liang et al., 2026].

5. Inference Strategy

RLDX-1 reduces real-robot inference latency by converting its forward pass into a single CUDA Graph and using workload-aware fused kernels. These optimizations target graph fragmentation, kernel-launch overhead, and memory traffic in its short-prefill execution pattern.

  • Motivation: Reducing per-step latency is critical because dynamic scenes can change while observations wait for inference and actuation, making resulting actions outdated.The robot operates in a closed loop of perception, inference, and actuation, so latency increases the mismatch between the observed state and executed action.
  • Static Graph Conversion: Static graph conversion moves rotary embeddings and attention-mask construction out of the forward pass, enabling single-graph capture and one launch per inference step.Under fixed deployment settings, these configuration-dependent values are precomputed and reused, eliminating subgraph fragmentation during CUDA Graph capture.
  • Short-Prefill Workload: RLDX-1’s short-prefill workload interleaves compute-bound matmuls with memory-bound operators, making operator grouping and data movement important for end-to-end latency.The full forward pass runs in one shot without autoregressive token generation, with relatively short sequence lengths across the VLM backbone and MSAT.
  • Workload-Aware Kernel Design: Workload-aware kernels combine operator fusion, on-chip memory reuse, and compute ordering beyond Torch Compile’s fixed graph-driven fusion path.The approach replaces inefficient operator groups identified through profiling while leaving the remaining implementation under Torch Compile.
  • Workload-Aware Kernel Design: Fusing operators reduces memory traffic by keeping intermediate tensors on-chip instead of repeatedly writing and rereading them between kernels.Without fusion, memory round-trips dominate runtime; fusion accesses memory once for input loading and once for output storage.

6. Evaluation and Analysis

RLDX-1 is evaluated across simulated and real-world robotic manipulation benchmarks against recent frontier VLAs. It consistently achieves leading performance, especially on challenging humanoid, motion-aware, and instruction-following tasks.

  • Simulation benchmarks: 97.8% on LIBERO, 81.5% on SIMPLER Google-VM, and 71.9% on SIMPLER WidowX establish RLDX-1 as the strongest model across the simulated benchmark suite.RLDX-1 consistently outperforms recent frontier models across the reported simulation benchmarks, including single-arm manipulation and sim-to-real visual transfer.
  • Simulation benchmarks: 58.7% on GR-1 Tabletop and 32.1% on RoboCasa365 exceed baselines, while RoboCasa Kitchen performance surpasses the 62–67% baseline range.The advantages are largest on challenging humanoid manipulation and diverse-task benchmarks, where most baselines remain below 50% on GR-1 Tabletop and average 14.8–26.9% on RoboCasa365.
  • Instruction-following tasks: 50% on Basic Pick-and-Place exceeds π0.5 (41.7%) and GR00T N1.6 (37.5%), while gains over π0.5 reach 14.6% and 12.5% on directional tasks.These results indicate stronger instruction-following performance, particularly for deformable-object manipulation under limited training data.
  • Motion-awareness tasks: RLDX-1 achieves the best performance across Conveyor Pick-and-Place tasks, while GR00T N1.6 averages 50.0% only at lower speeds and π0.5 averages 29.2% at the unseen speed.Baseline policies tend to follow fixed or previously seen conveyor speeds, revealing limited motion awareness and generalization.
  • Motion-awareness tasks: 97.9% on Spin Tracking and 81.5% on Pong Game substantially exceed π0.5 (32.3%) and GR00T N1.6 (26.0%) on motion-aware tasks.The baselines struggle to detect motion onset or respond across diverse directions, whereas RLDX-1 tracks and reacts to dynamic motion.

7. Related Work

Related work pursues generalist robotic policies through modular foundation-model approaches, synthetic data generation, and targeted capabilities for dynamic, long-horizon, contact-rich manipulation. These efforts address robotics’ limited data and the need for motion awareness, long-term memory, and physical sensing.

  • Foundation Models in Robotics: Foundation-model research seeks human-like generalist robotic policies that perceive, reason, and act across diverse tasks and environments, including through modular approaches.
  • Synthetic Data for Robot Learning: Synthetic-data research addresses robotics’ lack of internet-scale training data by generating robot trajectories in simulation through policies learned with imitation learning.
  • Functionalities for Robot Manipulation: Robot manipulation requires motion awareness, long-term memory, and physical sensing to handle dynamic, long-horizon, and contact-rich environments beyond vision alone.

8. Conclusion … Model Training Pipeline

RLDX-1 is presented as a general-purpose VLA for human-like dexterous manipulation, integrating motion awareness, long-term memory, and physical sensing for real-world deployment. The contributor sections document leadership and implementation across architecture, synthetic data, and model training.

  • 8. Conclusion: RLDX-1 targets human-like dexterous manipulation by unifying motion awareness, long-term memory, and physical sensing in a general-purpose VLA.The system also incorporates synthetic data augmentation for dexterous manipulation, although the supplied conclusion passage ends before detailing the full deployment design.
  • A. Contributors and Acknowledgments; A.1. Main Contributors; Project Leads: The project’s leads contributed across data and model training, data and architecture, model architecture, synthetic data, and model training.The listed project leads were Dongyoung Kim, Huiwon Jang, Myungkyu Koo, Suhyeok Jang, and Taeyoung Kim.
  • Research Leads: Dongyoung Kim, Huiwon Jang, and Jinwoo Shin led the overall research direction and coordinated work across contributor areas.The passage states that these research leads orchestrated cross-area coordination among contributors.
  • A.2. Core Contributors: Core contributors were grouped by area, with role-specific descriptions and repeated listings for members holding multiple roles.Names within each group were listed alphabetically by first name, and the grouping covered architecture, synthetic data, and model training pipeline work.
  • Model Architecture: Architecture contributors designed MSAT, the motion module, and physical sensory signal integration, while others supported early design and validation.The named contributions include early-stage MSAT design, motion-module implementation, physical-signal integration, and validation of that integration.
  • Synthetic Data: Synthetic-data contributors developed task and scene augmentation, ALLEX and GR-1 data generation, inverse-dynamics support, and motion-consistency or video-quality filtering.They also supported synthetic-data evaluation, real-world validation experiments, and video-model fine-tuning on ALLEX.
  • Synthetic Data; Model Training Pipeline: The contributor list identifies Junyoung Sung3 as supporting ALLEX synthetic-data generation and notes the KAIST-affiliated intern designation.These details appear in the synthetic-data contributor listing and its associated footnote.
  • Model Training Pipeline: The model-training pipeline team optimized multi-node VLM/VLA training and maintained the codebase through refactoring.Additional work supported baseline-model training, asynchronous inference, ALLEX deployment, real evaluation integration, and automated demo evaluation.

Real-World Robot Evaluation & Demonstration … B.1. Image Preprocessing

The reported work spans real-world evaluation, reinforcement learning, deployment optimization, infrastructure, data collection, benchmarking, release coordination, and image preprocessing. Together, these contributions support RLDX-1’s robot experiments, training and deployment pipelines, benchmark development, public release, and reduced vision-token usage.

  • Real-World Robot Evaluation & Demonstration: Real-world evaluation covered OpenArm, ALLEX, and FR3 experiments targeting motion, memory, physical sensing, and task-specific data refinement.Contributors supported OpenArm evaluation, designed motion-specific and memory-related ALLEX experiments, evaluated physical sensing on FR3 and ALLEX, and refined teleoperation data distributions for ALLEX tasks.
  • Reinforcement Learning: RECAP provided reinforcement-learning training, value modeling, inference functionality, and evaluation on the real ALLEX robot.The work included designing the RL task and value model, developing RECAP training and inference, and conducting real-robot ALLEX evaluations.
  • Inference Optimization & Deployment: Inference deployment combined asynchronous execution with static-graph conversion, CUDA Graphs, Triton and fused-attention kernels, and real-time chunking evaluation.The optimization pipeline addressed graph breaks and kernel performance while integrating and evaluating deployment components.
  • Training Infrastructure; Robot Control System: Training and robot-control infrastructure supported end-to-end experimentation through GPU and dataset management, teleoperation, low-level control, and fine-tuning data pipelines.The infrastructure covered training automation, robot interfaces, teleoperation control, and data collection for RB-Y1 and OpenArm.
  • Robot Hardware Research & Prototyping; Teleoperation System; Teleoperation Data Collection: Teleoperation systems and data collection supported ALLEX, OpenArm, Franka, and simulation workflows, including camera, video-streaming, VR, and automated evaluation components.The work included teleoperation backbones, operating environments, data collection across robots, and an automated simulation training-inference-evaluation pipeline.
  • Human Data Pipeline; Live Demo Production: Human-data pipelines established collection, processing, retargeting, training, and integration of human demonstrations, while live-demo teams collected data and fine-tuned RLDX-1.These efforts provided foundations for incorporating human demonstrations and preparing live demonstrations.
  • Dexterity Benchmark for Industry Conversion: DexBench defined an industry-oriented dexterity benchmark and standards for evaluating RLDX-1 in industry-conversion scenarios.The benchmark work included its curation, narrative and figure direction, and alignment of cross-team outputs.
  • Project Coordination & Public Release; Executive Sponsorship; B. Datasets Details; B.1. Image Preprocessing: Project coordination and public-release work managed technical narratives, demo production, licensing, and end-to-end release planning, while image preprocessing resized frames to at most 256^2 pixels using aspect-ratio-preserving 32-pixel multiples.The preprocessing procedure followed Qwen3-VL’s smart_resize approach to reduce the vision-token budget during training.

B.2. Public Real-World Real Data Details · B.3. Synthetic Data Details · C. Reinforcement Learning (RL) Details

RLDX-1 pretraining uses the OXE Magic Soup mixture, while synthetic trajectories are generated and filtered with embodiment-specific video models and inverse-dynamics action annotation. RL post-training combines a VLM critic with iterative advantage-based policy refinement and real-time chunked inference.

  • B.2. Public Real-World Real Data Details: RLDX-1 pretraining uses the OXE Magic Soup Open-X-Embodiment mixture, with its detailed subset composition reported in Table 5.The mixture follows the formulation used in Octo and OpenVLA.
  • B.3. Synthetic Data Details: Synthetic data uses video generative models fine-tuned on robot demonstrations to generate novel trajectories for GR-1 and ALLEX humanoid viewpoints and manipulation dynamics.Off-the-shelf video diffusion models are insufficient for egocentric robot perspectives and manipulation dynamics, motivating dedicated embodiment-specific models.
  • B.3. Synthetic Data Details: The GR-1 video model is fine-tuned on 3,027 ActionNet videos and 92 public NVIDIA GR-1 videos, with randomly sampled short, medium, or long Qwen3-VL captions.It is built on Cosmos-Predict2-14B and uses LoRA fine-tuning with rank 32.
  • B.3. Synthetic Data Details: The ALLEX video model mixes in-house ALLEX, OpenArm, and 3,027 ActionNet episodes at a 2:1:1 ratio to broaden manipulation-motion coverage.It is built on Cosmos-Predict2.5-2B and uses Qwen3-VL captioning for more detailed task descriptions.
  • B.3. Synthetic Data Details: Generated videos receive action labels from an inverse-dynamics model that predicts intermediate actions between observation frames using flow matching.The IDM uses a 0.1B Diffusion Transformer with a SigLIP-2 vision encoder.
  • C. Reinforcement Learning (RL) Details: RL post-training trains a Gemma3-4b-it VLM critic on target-task success demonstrations, then trains the policy for up to 30K steps with frozen vision features and 40-action chunks.The critic uses full-target LoRA with rank r = 128; policy training uses AdamW, learning rate 1 × 10−4, global batch size 128, and cosine decay with 5% warmup.

D. Parameter-Efficient Fine-Tuning … E.2. Implementation Details

The paper studies compute-efficient adaptation of RLDX-1 and complementary inference-time Best-of-N reasoning, detailing their evaluation settings, critic architecture, optimization, and sampling procedures.

  • D. Parameter-Efficient Fine-Tuning: LoRA is evaluated on the action model while fully fine-tuning the backbone VLM’s top-4 layers to assess whether performance can approach full fine-tuning with less compute.The supplied passage states that this configuration recovers full fine- but truncates the remainder of the result.
  • D. Parameter-Efficient Fine-Tuning: Table 6 compares full fine-tuning with three PEFT variants on Robocasa Kitchen using success rate, trainable parameters, and peak VRAM.Success rate averages 24 tasks × 50 episodes with three multi-frame views and 12 total frames; Full FT tunes the backbone VLM’s top-4 layers.
  • E. Test-time Sampling: Best-of-N sampling draws multiple action-chunk candidates and executes the candidate receiving the highest critic score, providing inference-time reasoning complementary to RL training.DEAS applies this pattern with an offline-trained chunk-level critic over VLA features.
  • E.1. Test-time Sampling Technique: At inference, RLDX-1 samples diversified candidate chunks with temperature 1.5–2.0 and selects the highest-Q chunk among N candidates.The VLA setup uses τ=0.7, γ1=0.9, γ2=0.99, action chunk length H + 1 = 16, AdamW learning rate 1×10^-4, batch size 64, and 30K steps.
  • E.2. Implementation Details: The critic uses a frozen post-RECAP1 RLDX-1 backbone, average-pools final-layer tokens, projects them to a shared 64-dimensional embedding, and combines this with proprioceptive state.Value and Q-functions are built on the resulting representation using an embodiment-conditioned 4-layer MLP.
  • E.2. Implementation Details: Critic training caches backbone features and optimizes only the critic with AdamW at 3×10^-5, using weight decay 10^-5, batch size 64, 10k steps, 5% warmup, and clipping 1.0.The learning rate was selected from a sweep over 1, 3×10^-5, and 10^-4.
  • E.2. Implementation Details: The RL critic follows DEAS with γ1=0.9, γ2=0.99, τ=0.7, and Polyak rate 0.005, while shifted rewards support a 101-atom HL-Gaussian distribution over [−100, 0].Scalar Q and V values are decoded by taking expectations over bin centers, with Gaussian smoothing σ=0.75 bin-widths.
  • E.2. Implementation Details: BoN inference rescales flow-matching noise by temperature T ∈[1.5, 2.0], with T=1.5 identified as the sweet spot, and shares one backbone pass across candidates.The implementation uses a B·N-tiled observation batch and noise sampled from N(0, T^2I).

E.3. Experimental results · F. Kernel Optimization · G. Evaluation & Analysis

RLDX-1’s test-time best-of-N sampling improves an incompletely converged policy but can degrade stronger checkpoints, while kernel optimization documents fused operators by comparing them with unfused sequences.

  • E.3. Experimental results: Best-of-N selection samples N candidate action chunks at temperature T, scores them with a learned Q-critic, and executes the highest-ranked chunk at each decision step.The method trades test-time compute for task performance and was calibrated through offline analysis of action diversity and Q values.
  • E.3. Experimental results: Offline analysis evaluates whether Q-based selection is effective by sampling N = 10 actions from RECAP3 and comparing their diversity and Q values.The analysis examines how temperature affects the gap between selected actions and how Q improvement varies across an episode.
  • E.3. Experimental results: The impact of test-time sampling depends strongly on policy convergence: it nearly raises RECAP1 to RECAP2 performance but harms already stronger RECAP2 and RECAP3 checkpoints.The same BoN configuration benefits the less-converged checkpoint while degrading later checkpoints.
  • G. Evaluation & Analysis: The supplied passages provide no specific experimental claim or result for the Evaluation & Analysis subsection.No passage assigned to subsection G reports an evaluation finding beyond the results summarized above.
  • E.3. Experimental results: 4.9 ± 1.3 attempts versus 8.5 ± 2.8 (−3.6) shows BoN sampling improves RECAP1 on Light Bulb Twisting, while RECAP2 and RECAP3 worsen by +2.3 and +2.2 attempts.The evaluated setting is N = 8 and T = 1.5; the result is shown across RECAP1–RECAP3 in Figure 23.
  • E.3. Experimental results: Increasing samples from 8 to 32 does not further improve RECAP1, indicating that the benefit comes from exploration rather than simply using more candidates.Figure 23 reports the absence of additional RECAP1 improvement when increasing the sample count.
  • F. Kernel Optimization: Kernel optimization reports the fused kernels used in the system and compares each fused operator sequence with its original unfused sequence.Table 7 provides the detailed kernel optimization results and identifies the input and output hidden states of layer ℓ.

G.1. Simulation Benchmark Details

RLDX-1 is evaluated across diverse simulation benchmarks spanning tabletop, kitchen, household, humanoid, robustness, and perturbation-based manipulation. Each benchmark uses task-specific datasets, camera configurations, evaluation episodes, and baseline protocols, with shared fine-tuning conventions.

  • Benchmarks: The simulation suite covers LIBERO, LIBERO-Plus, SIMPLER Google-VM/VA, SIMPLER WidowX, RoboCasa Kitchen, GR-1 Tabletop, and RoboCasa365.These benchmarks span single-arm tabletop and kitchen manipulation, robustness to diverse perturbations, humanoid tabletop tasks, and broader household manipulation.
  • LIBERO: LIBERO evaluates 40 tasks across Spatial, Object, Goal, and Long using fixed front and wrist cameras at 256×256 resolution and 50 trials per task.RLDX-1 is trained on concatenated training datasets from each sub-benchmark, with average success rates reported; LIBERO is built on a Franka Research 3 arm with a parallel gripper (Liu et al., 2023).
  • LIBERO-Plus: LIBERO-Plus measures robustness across object layout, viewpoint, initial state, language, lighting, background, and sensor-noise perturbations.Results for π0 and π0-FAST are taken from Fei et al. (2025), while remaining methods are evaluated on all 10,300 perturbation tasks.
  • SIMPLER: SIMPLER Google-VM/VA and WidowX each contain four tasks and use 200 evaluations per task with randomized benchmark seeds and average success rates.RLDX-1 is trained on Fractal for Google-VM/VA and BridgeV2 for WidowX, following the common setup in Li et al. (2024b).
  • RoboCasa Kitchen: RoboCasa Kitchen contains 24 mobile-Panda kitchen tasks and uses two fixed external cameras plus a wrist camera, all at 256×256 resolution.The tasks include pick-and-place, door and drawer operation, and appliance control; RLDX-1 is trained on concatenated machine-generated demonstrations.
  • GR-1 Tabletop: GR-1 Tabletop includes 24 humanoid tabletop tasks and reports average success over 50 episodes per task after training on 1,000 machine-generated demonstrations per task.The benchmark contains 18 object-rearrangement and 6 articulated-object manipulation tasks and uses a 256×256 ego-centric camera (Bjorck et al., 2025).

G.2. OpenArm Experiments Details · G.3. ALLEX Experimental Details · G.4. Franka Research 3 Experimental Details

The paper evaluates RLDX-1 on language-conditioned manipulation tasks across OpenArm, ALLEX, and Franka Research 3, using task-specific demonstrations, sensing modalities, hardware, and evaluation protocols. These experiments span tabletop pick-and-place, high-DoF humanoid manipulation, dynamic interactions, tactile sensing, and physical-signal control.

  • G.2. OpenArm Experiments Details: OpenArm evaluates language-conditioned pick-and-place in three-object tabletop scenes, using four object categories, three initial locations, and target placement instructions.The training categories are bottle, snack, cup, and doll, with distractors drawn from different categories.
  • G.2. OpenArm Experiments Details: OpenArm demonstrations cover Bottle to Shelf, Snack to Shelf, Cup to Dish Rack, and Doll to Box across Directional and Basic PnP families.Bottle to Shelf and Snack to Shelf are Directional PnP (Shelf); Cup to Dish Rack is Directional PnP (Dish Rack), while Doll to Box is Basic PnP.
  • G.2. OpenArm Experiments Details: OpenArm uses a two-arm OpenArm humanoid with 7-DoF arms, 6-DoF hands, a 2-DoF neck, and stereo egocentric vision collected through ExoArm-7 teleoperation.The setup uses a ZED 2i stereo camera mounted on the neck, with both left and right views as visual inputs.
  • G.3. ALLEX Experimental Details: ALLEX evaluates language-conditioned manipulation across task-specific environments, using 24 trials per task and varying conveyor speeds, boxes, desk heights, and other task conditions.Training excludes factors unrelated to the functional capability under evaluation, such as spatial generalization or language following.
  • G.3. ALLEX Experimental Details: ALLEX provides 48 manipulation joints, including dual 7-DoF arms, two 15-DoF hands, a 2-DoF waist, and a 2-DoF neck, with egocentric vision and torque signals from all joints.The benchmark uses only a head-mounted stereo camera and derives joint torques from motor currents.
  • G.4. Franka Research 3 Experimental Details: Franka Research 3 experiments evaluate language-conditioned manipulation with 24 trials per task, except Spin Tracking with 96 interactions and Pong Game with 54 interactions.Spin Tracking covers clockwise and counterclockwise motion across four initial positions and measures success rate.
  • G.4. Franka Research 3 Experimental Details: Franka Research 3 follows the DROID setup with a 7-DoF arm, Robotiq 2F-85 gripper, joint torque sensing, and a 15-dimensional AnySkin tactile signal.The tactile features come from five sensing units on the left gripper, each measuring three dimensions.

G.5. Additional ALLEX Experiments

On ALLEX’s bimanual Pot-to-Cup Pouring grasping task, co-finetuning RLDX-1 with synthetic data improved success across varied cup and coffeepot positions. Overall success rose from 66.7% to 83.3%, driven by fewer failures and more full successes.

  • Training Data: Training combined teleoperated real demonstrations with synthetic multi-turn rollouts generated by the ALLEX video generative model.Real demonstrations sampled cup and coffeepot positions from a Gaussian distribution around nominal locations, while synthetic generation reused preceding-turn final frames as conditioning.
  • Implementation Details: The co-finetuning setup mixed real and synthetic data equally and routed each dataset through its corresponding embodiment slot.RLDX-1 was fine-tuned for 30K steps with batch size 128, AdamW, learning rate 1 × 10^-4, 5% warmup, state dropout 0.4, and action horizon 40.
  • Results: Failures decreased from 7 to 2 while full successes increased from 9 to 14 after adding synthetic data.The improvement indicates enhanced spatial generalization and stable grasping across diverse cup and coffeepot positions.
  • Results: 83.3% overall success, up from 66.7%, was achieved by co-finetuning with synthetic data on 24 Pot-to-Cup Pouring grasping trials.Each trial scored full, partial, or failure; the evaluation covered three cup positions and four coffeepot positions with two trials per configuration.

G.6. Additional Simulation Results / Full Results

This section reports that larger training batch sizes improve RLDX-1’s simulation performance and provides full benchmark results across LIBERO, LIBERO-Plus, SIMPLER, RoboCasa Kitchen, GR-1 Tabletop, and RoboCasa365.

  • Effect of Batch Size: Larger training batch sizes improve RLDX-1’s performance on simulation benchmarks.This effect is reported in the batch-size comparison of Table 10.
  • Full Results: Full results are provided for RLDX-1 on LIBERO, LIBERO-Plus, SIMPLER, RoboCasa Kitchen, GR-1 Tabletop, and RoboCasa365 benchmarks.Tables 11–15 cover these benchmark results and include per-task success rates for the RoboCasa Kitchen, GR-1 Tabletop, and RoboCasa365 evaluations.
  • Full Results: 92.2% success is reported for Open Stand Mixer Head, while 0.0% is reported for Categorize Condiments, Gather Tableware, Heat Kebab Sandwich, Pan Transfer, and Separate Freezer Rack.The supplied per-task results also include 92.0% for PnP Toaster to Counter and 88.0% for Close Fridge.
Loading 2605.03269v2…