Skip to content
RLWRLDPublic

About

No description, website, or topics provided.

Resources

Stars

359 stars

Watchers

6 watching

Forks

Repository files navigation

RLDX-1.1

arXiv Project Page Models

RLDX-1.1 overview

RLDX-1.1 is a Vision-Language-Action model (VLA) for human-like dexterous manipulation. Beyond the versatile intelligence inherited from pre-trained VLM backbones, RLDX-1.1 adds three functional capabilities — motion awareness, long-term memory, and physical sensing — through a unified Multi-Stream Action Transformer (MSAT) architecture, a synthetic-augmented training pipeline, and a real-time inference stack.

Compared with RLDX-1, RLDX-1.1 scores higher on all four simulation benchmarks, reduces action jitter on low-DoF robots by up to 57%, and trains 2.27× and infers up to 3.7× faster. See What's New in RLDX-1.1.

Previous release: RLDX-1 (v1.0.2). All RLDX-1 checkpoints still load with this codebase.

News

  • [2026.11.12] RLDX-1.1 code and model checkpoints released! See What's New in RLDX-1.1.
  • [2026.06.16] RLDX-1-PT-IMG (lightweight image-input version) released!
  • [2026.05.06] Technical report, blog post, codes and model checkpoints for RLDX-1 released!

What's New in RLDX-1.1

Better Performance

We developed a stronger pre-training recipe for RLDX-1.1 by carefully refining the existing RLDX-1 corpus and incorporating a broad collection of open-source robot datasets. This improved data mixture provides a stronger foundation for downstream fine-tuning. To further improve the video-input model, we enable the motion module during fine-tuning, which was not used for RLDX-1 video-input checkpoints.

The table below compares RLDX-1.1 with RLDX-1 and recent VLAs across four public benchmarks. RLDX-1.1 improves on RLDX-1 on all four benchmarks, and an RLDX-1.1 model scores highest on each.

Comparison with Other VLAs

Success rates (%) of each model fine-tuned on each benchmark's training set. Note that some scores were re-evaluated with our current evaluation setup.

Method RoboCasa Kitchen GR-1 Tabletop RoboCasa365 (Avg) LIBERO (Avg)
π0.5 70.1 43.3 16.9 96.9
GR00T N1.7 64.8 40.2 24.6 95.1
RLDX-1 70.8 58.9 36.0 97.8
RLDX-1.1-IMG (ours) 70.4 60.6 33.2 97.2
RLDX-1.1 (ours) 72.3 60.3 40.2 97.9

Per-benchmark checkpoints, embodiment tags, and evaluation guides are listed under Reproducing Benchmark Results.

Smoother Actions in low DoF

For low-DoF embodiments, flow-matching noise in padded action dimensions can affect joint commands from the action DiT and increase jitter. RLDX-1.1 applies masking from pre-training to keep these unused dimensions at zero during training and inference.

We evaluated single-frame image models on a 13-DoF real robot with 50 actions per chunk. Values show mean direction changes in executed joint commands (reversals/s; lower is better).

Model Masking 4 denoising steps ↓ 10 denoising steps ↓
RLDX-1-IMG Off 9.69 10.38
RLDX-1.1-IMG On 6.45 4.47

Compared with RLDX-1-IMG, RLDX-1.1-IMG reduced reversals by 33% at four denoising steps and 57% at ten.

Each result averages metrics across 13 joints and five successful episodes at 30 Hz. The direction-change count excludes velocities with an absolute value of 0.06 rad/s or less.

This comparison includes other changes between model versions.

Keep mask_padded_action_dims=True during fine-tuning and inference with the resulting checkpoint.

Faster Training and Inference

RLDX-1.1 trains 2.27× faster than RLDX-1 on the same model and data, with the same training math. Every training optimization is on by default.

Inference is up to 3.7× faster on the RTX 5090, each release at its fastest setting on RoboCasa. It now runs batched and quantized from edge to cloud: fp8 on every supported GPU, plus nvfp4 on Jetson Thor, GB10 and the RTX 5090. RLDX-1 ran batch 1 in bf16, with optimized inference on the RTX 5090 only.

Hardware RLDX-1 RLDX-1.1
Training H200 × 8 338 samples/s 767 samples/s (2.27×)
Inference RTX 5090 75 ms 20 ms (3.7×)
Inference Jetson Thor not supported 86 ms

See Optimization for conditions, every preset on every device, and how to choose one.

Upgrading from RLDX-1

Checkpoints. Every RLDX-1 checkpoint loads with this codebase and keeps its outputs: a config.json without mask_padded_action_dims loads unmasked, and its 36 embodiment slots are kept. Checkpoints saved by this codebase use config.json format v2, which RLDX-1 cannot read.

Environment. Pinned dependencies changed (tyro, deepspeed, wandb and others), so run uv sync again. gymnasium moved to the eval extra.

Flags and APIs. Optional features and the data source are now subcommands (memory:on, data:paths, ...). A flag that does not apply to the chosen subcommands is rejected instead of silently ignored. Flags not listed below, such as --learning-rate and --n-cog-tokens, are unchanged; see the training flags.

RLDX-1 RLDX-1.1
--dataset-path(s) … --embodiment-tag … data:paths --data.paths … --data.embodiment-tag …
--dataset-mix-ratios … --data.mix-ratios …
--pt-dataset-root … --pt-dataset-mix … data:mix --data.root … --data.name …
--use-memory --memory-length 4 --concat-memory memory:on --memory.length 4 --memory.concat
--blockwise-attn-for-memory --memory.blockwise-attn
--use-motion --motion-insert-layer 9 motion:on --motion.insert-layer 9
--use-physics --physics-keys … --physics-dims … physics:on --physics.keys … --physics.dims …
--action-model-use-lora --action-model-lora-rank 16 action-model-lora:on --action-model-lora.rank 16
--backbone-use-lora --backbone-lora-rank 16 backbone-lora:on --backbone-lora.rank 16
--use-wandb --wandb-project … wandb:on --wandb.project …
--dataset-mode … Removed; data is always sharded
Training --rtc-inference-* Removed; RTC is chosen when serving
Server --compile … --runtime-profile … (default compile), see runtime profiles
Server --rtc-inference-mode … --rtc-inference-delay … rtc:trained or rtc:guided with --rtc.delay … --rtc.exec-horizon …
Server --dataset-path … --modality-config-path … (replay) rldx/eval/run_replay_server.py --dataset-path … --execution-horizon … --modality-config-path …
Server --strict, --use-sim-policy-wrapper, --verbose Removed
rollout_policy.py --model-path … Removed; start run_rldx_server.py and pass --policy-client-host … --policy-client-port …
rldx.policy.rldx_policy.RLDXPolicy(...) rldx.serving.policy.load_policy(...); get_action returns (actions, info)

Validation during training is new: pass --validation a.json b.json (one spec per file); see the training flags.

Behavior.

  • Run directories created by RLDX-1 cannot be resumed. Fine-tuning from a checkpoint into a new --output-dir is unaffected.
  • The policy server binds to 0.0.0.0 by default (RLDX-1: 127.0.0.1), so it listens on every interface; pass --host 127.0.0.1 for loopback only.
  • With RTC, the client sends the part of the previous chunk the robot has not executed yet as action_prefix. Memory checkpoints need session_ids. See the client protocol.

Model Architecture

The architecture is unchanged from RLDX-1.

RLDX-1.1 architecture
  • Multi-Stream Action Transformer (MSAT). Cognition, physics, and action each get a dedicated stream coupled by joint self-attention — an extension of MM-DiT to action modeling.
  • Motion awareness. Multi-frame observations + a motion module capture temporal dynamics; intermediate VLM layers compress video tokens to keep the policy efficient.
  • Long-term memory. A memory module fuses past cognition features with the current ones for history-grounded decisions beyond a short multi-frame window.
  • Physical sensing. Tactile and torque enter as a dedicated physics stream; the decoder is jointly trained to predict future physical signals.
  • Three-stage training. Pre-training (generalization) → mid-training (functionality) → post-training (task adaptation), with synthetic data augmenting rare manipulation scenarios.
  • Real-time inference. Inference runtime profiles combine regional torch.compile, one CUDA graph per call shape, and FP8 / NVFP4 quantization.

A five-stage walkthrough with config flags is in docs/architecture.md.

Model Checkpoints

Video input

Checkpoint Type Description
RLWRLD/RLDX-1.1-PT Base Pre-trained
RLWRLD/RLDX-1.1-FT-ROBOCASA Fine-tuned RoboCasa Kitchen
RLWRLD/RLDX-1.1-FT-GR1 Fine-tuned GR-1 Tabletop
RLWRLD/RLDX-1.1-FT-LIBERO Fine-tuned LIBERO
RLWRLD/RLDX-1.1-FT-RC365 Fine-tuned RoboCasa365

Image input

Checkpoint Type Description
RLWRLD/RLDX-1.1-PT-IMG Base Pre-trained
RLWRLD/RLDX-1.1-FT-ROBOCASA-IMG Fine-tuned RoboCasa Kitchen
RLWRLD/RLDX-1.1-FT-GR1-IMG Fine-tuned GR-1 Tabletop
RLWRLD/RLDX-1.1-FT-LIBERO-IMG Fine-tuned LIBERO
RLWRLD/RLDX-1.1-FT-RC365-IMG Fine-tuned RoboCasa365

All RLDX-1.1 checkpoints have 6.6B parameters.

RLDX-1 checkpoints (PT, PT-IMG, MT-*, FT-*) are listed in the RLDX-1 README and all load with this codebase.

Quick Start

Requirements: Python 3.10, CUDA 12.x, uv v0.8.4+

git clone https://github.com/RLWRLD/RLDX-1.git
cd RLDX-1
uv sync --python 3.10

Verify installation:

uv run python -c "import rldx; print(rldx.__version__)"

For simulator setup, dev tooling, and full troubleshooting, see docs/installation.md.

Load a checkpoint and predict an action chunk in-process:

from rldx.serving.policy import load_policy
from rldx.data.embodiment_tags import EmbodimentTag

policy = load_policy(
    model_path="RLWRLD/RLDX-1.1-FT-ROBOCASA",
    embodiment_tag=EmbodimentTag.GENERAL_EMBODIMENT,
    device="cuda:0",
)

# observation: {"video": {...}, "state": {...}, "language": {...}}
actions, info = policy.get_action(observation)

getting_started/01_quickstart.ipynb builds a dummy observation for this checkpoint and runs the call end to end.

Inference and Evaluation

Policy Server

rldx/eval/run_rldx_server.py serves the same policy over ZeroMQ for simulator rollouts and real-robot deployment:

uv run python rldx/eval/run_rldx_server.py \
    --model-path RLWRLD/RLDX-1.1-FT-ROBOCASA \
    --embodiment-tag GENERAL_EMBODIMENT \
    --host 0.0.0.0 --port 20000

Two orthogonal knobs control real-time behavior:

  • --runtime-profile — a preset per recipe: compile (bf16, default), eager (bf16, no compile), any:bf16 or any:fp8 on any device, and <device>:nvfp4-mix (text fp8 except gate_up_proj in nvfp4) or <device>:nvfp4 for thor and gb10. Presets other than compile and eager turn on forward CUDA graphs and need limits:on --limits.batch B --limits.instruction-tokens T, which bounds requests.
  • rtc:guided / rtc:trained — Real-Time Chunking for chunk-boundary stitching. Off when omitted.

The client protocol, full flag list, and a walkthrough of the trade-offs are in docs/inference_server.md.

Reproducing Benchmark Results

Each benchmark has a self-contained eval README; this table maps each column in Comparison with Other VLAs to the fine-tuned checkpoint, the embodiment tag the server expects, and the runnable guide.

Benchmark Fine-tuned Checkpoint Embodiment Tag Eval Guide
LIBERO RLWRLD/RLDX-1.1-FT-LIBERO GENERAL_EMBODIMENT run_scripts/eval/libero/README.md
RoboCasa Kitchen (24 tasks) RLWRLD/RLDX-1.1-FT-ROBOCASA GENERAL_EMBODIMENT run_scripts/eval/robocasa_kitchen/README.md
GR-1 Tabletop RLWRLD/RLDX-1.1-FT-GR1 GENERAL_EMBODIMENT run_scripts/eval/gr1_tabletop/README.md
RoboCasa365 RLWRLD/RLDX-1.1-FT-RC365 GENERAL_EMBODIMENT run_scripts/eval/robocasa_365/README.md

Shared mechanics (server + rollout split, common flags, troubleshooting) are documented in docs/evaluation.md.

Fine-tuning

This section covers how to fine-tune RLDX-1.1 from a pre-trained checkpoint (RLWRLD/RLDX-1.1-PT) on your own LeRobot v2.1 dataset. The training entry point is a single CLI (rldx/experiment/launch_train.py) where flags toggle the optional functional capabilities described in Model Architecture:

  • --video-length N — temporal frames per observation (motion awareness)
  • memory:on — temporal memory module (long-term memory)
  • motion:on — motion module inside the VLM backbone
  • physics:on --physics.keys ... — tactile / torque streams (physical sensing)

LoRA, training-time RTC, and the full flag list are documented in docs/training.md. Below are the canonical recipes.

Data Preparation

RLDX-1.1 uses LeRobot v2.1 format datasets. run_scripts/data/ has a converter from HDF5 teleoperation recordings (convert_h5_to_v2.py) and wrappers for one dataset or a directory of datasets; the scripts define the expected input layout and arguments. The converter needs LeRobot 0.3.3, the last release that writes v2.1 (uv pip install --no-deps lerobot==0.3.3 jsonlines h5py).

Each dataset must carry a meta/modality.json that slices the flat state / action vectors into named joint groups and remaps video columns to modality keys — see docs/training.md. For a custom robot, pass your modality config with --modality-config-path together with an EmbodimentTag that selects the per-robot MLP head slot (default: GENERAL_EMBODIMENT; see docs/embodiment_tags.md). The EmbodimentTag design follows NVIDIA GR00T N1.7.

Basic Recipe

uv run torchrun --nproc_per_node=<N_GPUS> rldx/experiment/launch_train.py \
    --base-model-path RLWRLD/RLDX-1.1-PT \
    data:paths --data.paths /path/to/your/dataset \
    --data.embodiment-tag GENERAL_EMBODIMENT \
    --modality-config-path /path/to/your_modality_config.py \
    --video-length 4 \
    --n-cog-tokens 64 \
    --num-gpus <N_GPUS> \
    --global-batch-size 64 \
    --learning-rate 1e-4 \
    --max-steps 60000 \
    --save-steps 5000 \
    --output-dir ./outputs/my_finetune

--modality-config-path is required: it registers the modality config for the embodiment tag (see Data Preparation).

With Memory, Motion, and Physics

Recommended for embodiments where memory, motion awareness, or contact sensing matter. To enable a single add-on instead of all three, keep just the corresponding memory:on / motion:on / physics:on group(s) and drop the rest.

uv run torchrun --nproc_per_node=<N_GPUS> rldx/experiment/launch_train.py \
    --base-model-path RLWRLD/RLDX-1.1-PT \
    data:paths --data.paths /path/to/your/dataset \
    --data.embodiment-tag GENERAL_EMBODIMENT \
    --modality-config-path /path/to/your_modality_config.py \
    --video-length 4 \
    memory:on --memory.length 4 --memory.concat \
    motion:on --motion.insert-layer 9 \
    physics:on --physics.keys tactile torque --physics.dims 30 7 \
    --new-param-warmup-steps 2000 \
    --n-cog-tokens 64 \
    --num-gpus <N_GPUS> \
    --global-batch-size 64 \
    --max-steps 60000 \
    --output-dir ./outputs/my_finetune_all

Key Training Flags

Flag Description Default
--video-length Number of video frames (video token compression is always on; set to 1 for single-frame) 4
--video-stride Stride between frames in action-step units 2
memory:on Enable temporal memory module off
--memory.length Memory context window (timesteps) 4
motion:on Enable motion module off
physics:on Enable physics signal conditioning off
--n-cog-tokens Number of cognition tokens 64
--global-batch-size Total batch size across GPUs per step, split evenly over --num-gpus 64
--new-param-warmup-steps Warmup steps for newly added modules 0

LoRA

For memory-constrained fine-tunes you can replace full-parameter tuning of the action model (MSAT) and/or the backbone VLM with PEFT LoRA adapters:

action-model-lora:on --action-model-lora.rank 16 --action-model-lora.alpha 32
backbone-lora:on --backbone-lora.rank 16 --backbone-lora.alpha 32 --backbone-lora.num-layers -1

Only the adapters train in the part a LoRA covers, so action-model-lora:on does not take --tune-diffusion-model, and backbone-lora:on does not take --tune-llm, --tune-visual, --tune-top-llm-layers or --freeze-cog-tokens; giving one is an error. Full flag list and target-module defaults are in docs/training.md.

Training for Real-Time Chunking

If you intend to serve the checkpoint with rtc:trained (faster, fullgraph-compatible), enable training-time RTC at training time:

--rtc-training-max-delay 4

The training-time RTC formulation follows Black et al. (Training-Time Action Conditioning for Efficient Real-Time Chunking); the inference-side counterpart is Black et al. (Real-Time Execution of Action Chunking Flow Policies). See docs/training.md and docs/inference_server.md for usage details.

Documentation

Hands-on guides live under docs/:

Guide What it covers
installation.md Environment setup, simulator venvs, dev tooling, common pitfalls
architecture.md Five-stage walkthrough of the RLDX-1.1 model and its config flags
training.md launch_train.py recipes (fine-tune / mid-train), LoRA, training-time RTC, dataset layout
embodiment_tags.md What EmbodimentTag is and how to pick one for a custom robot
evaluation.md RoboCasa / LIBERO / GR-1 eval, server + rollout split, results aggregation
inference_server.md run_rldx_server.py CLI, wire protocol, RTC modes, runtime profiles, simulator + real-robot deployment
optimization.md Training speedups and what they change for your runs, inference execution path, fp8 / nvfp4 quantization, choosing a runtime profile

Step-by-step notebooks are in getting_started/.

Acknowledgments

RLDX-1.1 builds upon the following open-source projects:

Contributing

We currently do not accept external pull requests on this repository. If you encounter a bug, broken reproduction step, or have a question about RLDX-1.1, please open an issue at github.com/RLWRLD/RLDX-1/issues and we will follow up there.

License

  • Code: released under the Apache License 2.0. The codebase is built on the NVIDIA Isaac GR00T N1.7 framework. Third-party code included in this repository is listed, with its copyright notices and license texts, in NOTICE.
  • Model weights: distributed on Hugging Face under the RLWRLD Model License v2.0 (a non-commercial license with attribution and share-alike terms). By using any RLWRLD/RLDX-1.1-* checkpoint you agree to those terms. RLWRLD/RLDX-1-* checkpoints remain under the RLWRLD Model License v1.0.

Citation

@article{rldx2026,
  title={RLDX-1 Technical Report},
  author={Dongyoung Kim and Huiwon Jang and Myungkyu Koo and Suhyeok Jang and Taeyoung Kim and others},
  year={2026},
  journal={arXiv preprint arXiv:2605.03269},
  eprint={2605.03269},
  archivePrefix={arXiv}
}

About

No description, website, or topics provided.

Resources

Stars

359 stars

Watchers

6 watching

Forks

Releases

Packages

Contributors

Languages