Ansys Action Potential Model

3Eth Ai Center

https://cvlab-kaist.github.io/Geometric-Action-Model

Geometric Foundation Model

...

Layer

Layer

...

Instruction

...

T+1 T+2

...

Feature

Figure 1: GAM repurposes geometric foundation models into fast and robust robot policies. (a) GAM jointly predicts future 3D geometry and action chunks within a shared geometric backbone. (b) By leveraging explicit 3D geometric priors, GAM improves robustness and real-world perfor- mance while reducing latency and model size compared to existing baselines.

Abstract: Generalist robot policies must follow user instructions while reasoning about how objects, cameras, and robot actions interact in the 3D physical world. Recent vision-language-action models (VLAs) and video world-action models (WAMs) inherit strong semantic or temporal priors from large-scale foundation models, but they still operate primarily on 2D image frames or 2D-derived la- tent spaces, leaving implicit the 3D geometry required for contact-rich manipula- tion. We propose the Geometric Action Model (GAM), a language-conditioned manipulation policy that directly repurposes a pretrained geometric foundation model (GFM) as a shared substrate for perception, temporal prediction, and ac- tion decoding. GAM splits the GFM at an intermediate layer: the shallow layers serve as an observation encoder, and a causal future predictor inserted at the split layer forecasts future latent tokens conditioned on language, proprioception, and action history. The predicted future tokens are then routed through the remain- ing GFM blocks for feature propagation and decoding, allowing a single back- bone to produce both future geometry and actions. This design equips the GFM with language-conditioned temporal world modeling through minimal architec- tural modification while preserving its rich geometric priors. Across a broad suite of simulation and real-robot manipulation benchmarks, GAM is more accurate, more robust, faster, and lighter than current foundation-model-scale baselines.

∗Equal contribution. †Corresponding author.

Introduction

A long-standing goal in robotics is to build generalist manipulation policies that can follow natural- language instructions and manipulate arbitrary objects across diverse scenes [1, 2, 3, 4]. To achieve this, a general manipulation model must not only recognize objects and parse instructions, but also reason about how the physical world will evolve under its own actions. This requires a unified understanding of language, visual appearance, scene geometry, robot state, and physical dynamics.

Recent progress has therefore increasingly relied on large-scale foundation models as pretrained substrates for robot policies. Vision-language-action models (VLAs) build on vision-language mod- els whose representations are aligned with natural language, and learn to map visual and linguistic tokens to robot actions [5, 1, 2, 4, 6, 7]. Video world-action models (WAMs) instead leverage pre- trained video generation models, using their world prediction priors to jointly model future frames and actions [3, 8]. While these approaches have shown impressive language-conditioned manipu- lation ability, they are fundamentally in 2D: 3D cues such as depth, scale, and occlusion are left implicit in monocular cues that the action decoder must disentangle on its own, leading to lim- ited generalization across environment changes, especially in robot initial state and camera view- point [9, 10].

To overcome this limitation, recent work incorporates 3D geometric information into robot poli- cies. One line learns policies directly on explicit 3D observations such as raw point clouds [11, 12], demonstrating the value of geometry for generalization but typically requiring task-specific encoders trained from scratch. With the emergence of Geometric Foundation Models (GFMs) [13, 14, 15], some works transfer pretrained geometric priors into VLA policies, either distilling selected GFM features into the VLA backbone through representation alignment [16, 17, 18] or attaching a lightweight action head on top of a GFM’s final features [19, 20]. These improve spatial aware- ness, but use the GFM only as a static feature extractor: its multi-layer geometric structure is never repurposed as the policy’s own temporal and action-generating substrate.

In this work, we propose Geometric Action Model (GAM), which directly repurposes a GFM as a manipulation policy by using it as a shared medium for perception, future-state prediction, and action decoding. We show that by jointly predicting future action and geometry, geometric world dynamics can be inherently incorporated into robot policies.

Specifically, we split the pretrained GFM at an intermediate layer: the shallow layers serve as an observation encoder, while the remaining layers serve as a decoder block. Given the current visual observation, the observation encoder extracts spatially meaningful scene representations. To model how the world evolves over time, we insert a causal transformer at the intermediate layer that predicts future feature representations. This predictor is conditioned on task language, proprioception, and action history by introducing them as additional tokens at each timestep. The predicted future-state tokens are then processed by the remaining GFM decoder together with an action token, allowing the backbone to produce both future geometry and robot actions. An intuitive comparison between existing paradigms and our proposed framework is illustrated in Figure 2.

Across diverse simulation [21, 9, 22] and real-world benchmarks, GAM matches or exceeds the success rate of current foundation-model-scale baselines such as VLAs and WAMs while using substantially fewer trainable parameters and substantially faster inference (55× faster), and shows improved generalization to unseen scenarios. GAM especially achieves outstanding performance in camera perturbation settings (↑9.7%p), which requires geometric understanding priors.

Our Contributions Are As Follows:

• We introduce Geometric Action Model (GAM), a manipulation policy that combines tem- poral world modeling, latent feature-space prediction, and a geometric foundation-model substrate in a single shared-backbone architecture.

• We show that action and geometry can be predicted in a shared token space: a single autoregressive sequence and a single backbone forward pass produce both action tokens and future-scene tokens, decoded by a lightweight action regression head and a depth head.

Action Head

Future Geo.

(C) Gam

Figure 2: (a) Video WAMs [3, 8, 23] operate in 2D pixel space, predicting future latents and actions via video diffusion. (b) Geometry-aware VLAs [17, 18] predict actions using a VLA with passive feature distillation from an external GFM. (c) GAM (ours) unifies perception, geometry prediction, and action decoding by inserting a geometric world model inside a single GFM.

• We demonstrate that across diverse simulation and real-world manipulation benchmarks, GAM is simultaneously more accurate, more robust, faster, and lighter than current foundation-model-scale alternatives.

Related Work

Vision-Language-Action Models. Vision-language-action models (VLAs) adapt a pretrained vision-language model (VLM) into a robot policy. Early works [5, 1] autoregressively decode dis- crete action tokens from a finetuned VLM. This paradigm is extended through parallel decoding , flow-matching action experts [4, 6, 24], diffusion-based action heads , frequency-space action tokenizations , and compact open-source VLMs [26, 27]. This line of work extends large general- ist foundation models for humanoid and embodied control [28, 29, 30, 31], to spatial-representation variants , and to refinements in training and post-training [33, 34, 35]. While these VLAs lever- age strong open-vocabulary recognition to produce policies, they rely solely on 2D image priors and therefore lack 3D understanding. To address this, recent works [17, 18, 36] attempt to align intermediate VLA features with geometric foundation model. However, they do not fully exploit the geometric understanding of geometric foundation model backbone.

World Action Models for Robot Manipulation. World action models are trained to predict future states in order to learn a policy. One branch builds on large pretrained video generation models , fine-tuning them to jointly predict future frames and actions [3, 8, 23], arguing that the benefit of such co-training stems primarily from training-time supervision rather than test-time future imag- ination [23, 7]. A second branch keeps the visual backbone frozen and trains a separate temporal predictor in its feature space for planning via model-predictive control [38, 39], with related work using implicit future-latent alignment as an auxiliary training signal . However, these video backbones encode only 2D image-space priors, which do not explicitly resolve depth, scale, or oc- clusion. GAM instead predicts in the latent space of a geometric foundation model that encodes rich 3D priors, while inheriting from both branches the paradigm of using future prediction as a training signal.

Geometric Foundation Models for Manipulation. Geometric foundation models (GFMs) [14, 13, 41, 42] infer dense 3D structures from multi-view images and have recently served as perceptual substrates for robot policies. Early approaches integrate GFMs with VLAs as frozen feature ex- tractors via representation alignment [17, 18], direct encoder replacement [36, 19], or point-cloud fusion . Moving beyond static extraction, recent concurrent works adopt GFMs for predictive control: Song et al. utilizes the GFM to jointly predict actions and current-frame 3D properties, while Xu et al. employs the GFM with a diffusion policy to co-denoise future action chunks and

Causal Transformer Layer

... ... ...

Text

Enc.

Geometry Head

...

State

Prev.

Action

Geo. Feat.

Gfm Deeper Layers

... ...

State

Prev.

Action

Geo. Feat.

(B) Block-Causal Attention Mask

Figure 3: Main architecture of GAM. 3D latents. GAM departs from these paradigms in two ways: (1) action and future-scene predictions are jointly modeled within a single autoregressive token sequence rather than separated heads or diffusion processes, and (2) the GFM’s deep blocks are explicitly repurposed to decode predicted future tokens rather than merely processing observed ones.

Preliminaries: Geometric Foundation Models

A geometric foundation model (GFM) such as VGGT or DA3 is a feed-forward transformer that maps one or more RGB images to dense 3D geometry. Given a sequence of V views I =

{Iv}V

v=1 with Iv ∈R3×h×w, a GFM produces per-pixel depth Dv ∈Rh×w or 3D point maps Pv ∈R3×h×w in a shared world frame, and per-view camera intrinsics and extrinsics (Kv, ξv) ∈ R3×3 × SE(3) via auxiliary heads. Specifically, each view Iv is partitioned into P non-overlapping patches of size p × p and projected by a patch embedding into a per-view token sequence:

V}P

j=1 are patch tokens, and d is the hidden dimension. The full input sequence concatenated across views is Z(0) = [z(0)

1 , . . . , Z(0)

V ] ∈RV (1+P )×d. These tokens are processed by a stack of M transformer blocks {f (m)}M

Attention Modes. Frame-Wise Attention F (M)

frame operates within each view tokens independently, at- tending over the (1 + P) tokens of a single image. Global attention f (m)

Global Operates Jointly Over All

V (1 + P) tokens, fusing information across viewpoints. The hidden state at the m-th transformer

Frame, F (M)

global}.

(2)

After the transformer forward pass, dense geometry is decoded from multiple intermediate hidden states Z(m∗), where m∗is one of the selected layers in S = {m1, m2, m3, m4}. The extracted multi-layer hidden states are then fed into the DPT head to estimate per-pixel geometry.

Gam: Geometric Action Model

Problem Formulation. We consider language-conditioned robot manipulation. At each timestep t, the robot receives a multi-view RGB observation ot = {Iv,t}V

V=1 From V Fixed Cameras, A Pro-

prioceptive state st ∈Rds describing the robot’s joint configuration and end-effector pose, and a natural-language task instruction ℓthat is held constant throughout an episode. The policy πθ must produce an action chunk ˆat ∈RC×da of length C, encoding the next C delta-pose or joint com- mands to be executed open-loop before the next observation is acquired. We learn a policy:

Πθ :

{ot−H+1, . . . , ot}, {st−H+1, . . . , st}, {at−H, . . . , at−1}, ℓ

4

from a dataset of N expert demonstrations D = {(τi, ℓi)}N

(Ot, St, At)Ti

t=1 pairs a sequence of observations, states, and executed action chunks with a fixed instruction ℓi. The policy conditions on a context window of H recent timesteps. Overview.

In the following sections, we explain how we transform a pretrained GFM into a language-conditioned world-action model. Our key idea is to split the GFM into two parts and insert a causal temporal predictor between them. This design lets GAM formulate future prediction directly inside the GFM latent space, enabling all predictive computation to be performed in the GFM’s geometric representation space.

Concretely, our framework operates in three sequential stages inside the GFM. First, the observation encoder (§4.1) repurposes the shallow layers of the GFM to extract latent geometric features from multi-view observations. Next, the causal future predictor (§4.2) operates at the split layer, where it combines these geometric features with language, proprioception, and action history to predict future latent tokens. Finally, during feature propagation and decoding (§4.3), the predicted future tokens are routed through the remaining deep GFM blocks to simultaneously decode future geometry and the final action chunk ˆat. Figure 3 (a) shows the overall architecture of our model.

Observation Encoder

We first reuse the shallow layers of the pretrained GFM as the observation encoder. Let Ls denote the split layer where the causal future predictor is inserted. The original GFM transformer stack is

E≤Ls = F (Ls) ◦· · · ◦F (1),

D>Ls = f (M) ◦· · · ◦f (Ls+1).

(4)

Here, the choice of Ls is important because Ls must be deep enough to extract sufficiently rich visual features from the raw observations, yet shallower than the earliest layer used in the DPT head Ls < m1, so that predicted future states can be decoded into future geometries by the DPT heads.

After defining this split layer Ls, for each timestep t′ in the context window, we tokenize the multi-

View Rgb Observation Ot′ = {Iv,T′}V

v=1 using the original GFM patch embedding. This produces

(5)

where each view contributes one camera token and P patch tokens. The observation encoder maps these tokens to the split-layer representation Z(Ls)

. By Applying This Encoding Independently To Each

timestep in the context window, the output of this stage is a sequence of per-timestep geometric latent

T

}.

Causal Future Predictor

After the observation encoder, GAM performs temporal prediction directly at the split layer Ls, forecasting the next latent geometric state from current and past observations while conditioning on the task instruction, proprioception, and action history. To this end, we insert a causal future predictor gϕ between the shallow encoder E≤Ls and the deep decoder D>Ls. For each timestep t′ in the context window, the encoder provides latent tokens Z(Ls)

(6)

with ψs, ψa lightweight projection layers, and the instruction ℓinto language tokens Lℓwith a pretrained text encoder. We then form a per-timestep token block by concatenating the encoded GFM tokens with the proprioception and action-history tokens Ut′ = [pt′; qt′; Z(Ls)

]. The Full

input to the causal future predictor is X = [Lℓ; Ut′−H+1; . . . ; Ut′]. The combined sequence X is then processed through block-causal self-attention , ensuring the model incorporates past and present contexts without future leakage, as illustrated in Figure 3 (b). At

5

the final layer of the predictor gϕ, we read off the predictions from their respective sequence slots. Specifically, the hidden states corresponding to the geometry slots forecast the latent geometric

Tokens Of The Future Frame, Denoted As ˜Z(Ls)

t′+1. Concurrently, the hidden state of the designated previous-action slot is projected to produce a predicted next action token ˜at′ ∈Rd, in direct analogy to next-token prediction in a causal language model. By jointly forecasting action and geometric latents in this layer, we ensure that action tightly interacts with spatial representations.

This design of introducing a causal transformer predictor gϕ allows the pretrained GFM to acquire language-conditioned temporal world modeling with minimal architectural modification. Only the inserted gϕ needs to learn how to fuse language, proprioception, and action history with GFM latent

Features. The Resulting Predictions, Ez(Ls)

t′+1 and ˜at′, are then passed to the remaining GFM blocks for joint geometry and action decoding.

Feature Propagation And Action Decoding

Following the causal future predictor, the single action token ˜at′ is replicated V times to form a set

Of Per-View Action Tokens {˜Av,T′}V

v=1, where ˜av,t′ = ˜at′. Concatenated with the geometry tokens, they are fed through the remaining GFM blocks D>Ls. We perform this feature propagation by appending each view’s corresponding action token ˜av,t′ directly to its geometry token sequence for

V,T′+1; ˜Av,T′I

.

(7)

To prevent future leakage, we extend the predictor’s causal mask strategy to the GFM’s remaining

Global Attention Layers (F (M)

global). Finally, the propagated features are decoded by two heads. The lightweight action head hact ag- gregates action tokens over the context window to regress the executable action chunk ˆat′, while the original GFM depth head hdepth decodes geometry tokens into action-aligned future depth maps.

The GFM’s deep blocks, originally pretrained to decode shallow features into 3D geometry, are thus repurposed here as the decoder of the world model’s predicted future.

Training And Inference

The policy is trained end-to-end by minimizing a multi-task objective over action execution, world

(8)

where the λ factors balance each term and H = {t−H+1, . . . , t} is the context window. The action loss Lact is an ℓ1 regression between the decoded action chunk ˆat′ and the expert action at′ over all t′ ∈H. The future-feature loss Lfeat anchors the predictor gϕ to temporal geometric transitions by

T′+1

1 .

(9)

The future-depth loss Ldepth grounds the predicted future in valid 3D structure by supervising the

Decoded Depth ˜Dt′+1 = Hdepth(˜Z(M∗)

t′+1 ) using depth head hdepth against ground-truth future depth Dt′+1, adopting the scale-invariant and gradient-matching penalties of the GFM [13, 15]. At inference, we maintain the historical context online with key-value caching, so each step pro- cesses only the new observation ot and previous action at−1 in a single feed-forward pass.

Implementation Details

We use DA3-Giant fine-tuned on Track4World as the backbone. We insert a 12-layer causal predictor with width dg = 1024 at layer Ls = 12, where alternating attention begins. For

6

Table 1: Evaluation results on LIBERO and LIBERO-Plus. Success rates are reported in % with absolute performance drops from LIBERO to LIBERO-Plus shown in parentheses. Color highlights denote the top three performing methods within each column: first , second , and third .

Size

Orig.

Plus

Cam.

Robot

Lang.

79.1

the task instruction, we extract language tokens using a frozen T5 encoder . The policy uses a context horizon of H = 4 for pre-training and H = 1 for post-training and predicts C = 8 step action chunks in a da = 7 end-effector action space from ds = 7 proprioceptive states. We pretrain GAM on 784K single-arm robot trajectories from RoboCasa365 , MimicGen , and OpenX- Embodiment , then post-train it on each benchmark. We optimize with AdamW using a constant learning rate, freeze layers before Ls and the depth head, and supervise depth with simulator ground truth. We set λact = 3, λfeat = 1, and λdepth = 3. Further details are provided in the appendix.

Experimental Setup

Simulation Benchmarks. We evaluate generalization across distinct axes using two simulation benchmarks. Specifically, we train our policy on LIBERO , a lifelong single-arm manipula- tion benchmark spanning diverse spatial layouts, object identities, and task goals. To rigorously assess out-of-distribution robustness, we then evaluate the trained models in a zero-shot manner on LIBERO-Plus , which introduces controlled environmental perturbations across dimensions such as camera viewpoint, lighting, and backgrounds. We report additional results in the appendix.

Real-Robot Setup. We train on four manipulation tasks (∼200 demonstrations each) using wrist- mounted and third-person cameras, adhering to the simulation protocol. Since ground-truth geome- try is unavailable in the real world, target future depth maps are obtained as pseudo-labels directly from the pretrained backbone GFM. We evaluate robustness via 20 trials per task, divided equally between nominal setups and perturbed environments, specifically varying external camera positions.

See the appendix for robot environment with full task and evaluation details. Baselines. We compare GAM against representative baselines from three families discussed in §2: VLAs [1, 2, 4, 6, 7, 26, 33, 34], WAMs [3, 35, 52], and geometry-aware VLAs [17, 18]. For the real-robot setup, we compare against π0.5 and Spatial Forcing . For fairness, comparisons utilize a matched evaluation protocol, with performance numbers either re-evaluated using available checkpoints or taken directly from their respective published benchmarks.

Covered Pot

Figure 4: Real-world robot tasks and results. Each task is evaluated under both in-domain (Light bar) and out-of-domain (Dark bar) settings. The illustration of each task is shown on the right.

Main Results

Simulation Results. As shown in Table 1, GAM achieves highly competitive success rates on the standard LIBERO benchmark, where performance is heavily saturated. Crucially, on the more challenging LIBERO-Plus benchmark, our model consistently outperforms competing baselines, demonstrating a remarkable improvement in the camera-perturbation setting (↑9.7%p). This gain highlights the advantage of our end-to-end integration of the GFM. While existing geometry-aware VLAs only partially exploit GFM representations, GAM embeds the GFM throughout its entire predictive pathway to yield a deeply geometry-aware policy.

Real-world Results. To examine whether the gains observed in simulation transfer to physical execution, we additionally evaluate GAM in a real-world setting. Figure 4 shows that GAM sub- stantially outperforms all baselines. In particular, our model remains robust under out-of-domain conditions (the camera-perturbation setting) where other baselines struggle. These results demon- strate that GAM generalizes to the real-world domain and is robust under perturbations, owing to its thorough exploitation of the GFM when training the policy.

Ablation Study

Post-training Component Analysis. Table 2 summarizes ablation study of key post-training com- ponents on Object suite of LIBERO and LIBERO-Plus. Pretraining is crucial for robustness: omit- ting it mildly affects nominal LIBERO but severely degrades LIBERO-Plus. With a pretrained backbone, removing Ldepth or Lfeat has minimal impact, suggesting geometric dynamics are already encoded. Notably, even without pretraining, these future-prediction losses provide strong geometric supervision and substantially improve robustness on LIBERO-Plus. Finally, the horizon ablation shows that H = 1 is sufficient and more robust than longer histories, consistent with prior observa- tions that extended context can introduce spurious correlations [53, 54].

Table 2: Component ablation.

H

Orig.

50.0

Table 3: Layer ablation.

0.0

0.0

Table 4: Inference cost.

Split Layer Ls And Re-Initializing

the predictor.

Could Isolate The Effect Of Lfeat

itself.

Mance, Validating It As The Opti-

mal seam between frame-wise and cross-view attention. While layer 19 remains competitive, insert- ing the predictor too early (Ls = 0) or late (Ls ∈{27, 33, 39}) causes total performance collapse. This confirms that forecasted tokens require sufficient interaction through deep layers to properly integrate into the pretrained 3D geometric prior.

Figure 5: Success Rate Vs. Cam-

era perturbation difficulty. Inference Speed and Model Size. As shown in Table 4, GAM achieves the lowest latency among all baselines, requiring only 6.9 ms (≈145 Hz) for a single feed-forward pass and running up to 55× faster than the diffusion-based Cosmos Policy.

All methods are benchmarked under the same setup, with further details provided in the appendix. By utilizing single-pass predic- tion, GAM avoids the multi-step denoising of diffusion policies, achieving low latency while matching prior accuracy and robust- ness with only 1.4B parameters.

Robustness to Viewpoint and Scene Variation. Figure 5 further breaks down the camera-perturbation results by difficulty level of LIBERO-Plus . GAM achieves consistently higher success rates than all baselines at every level, and the advantage remains clear even under the strongest perturbations.

Conclusion And Limitation

We introduced Geometric Action Model, which unifies geometry and action prediction with tem- poral world modeling inside a single shared GFM. By inserting a causal transformer between the GFM’s shallow and deep layers, GAM autoregressively decodes actions and future geometries, re- solving the spatial ambiguities of traditional foundation-model substrates. Across extensive sim- ulation and real-world benchmarks, GAM achieves superior accuracy, faster inference, and strong out-of-distribution robustness to environmental perturbations. The framework also has limitations.

Its language reasoning and commonsense capabilities are bounded by the frozen text encoder; inte- grating a large language model or an external reasoning module is a natural next step.

Acknowledgments

This work was supported under project ID a144 as part of the Swiss AI Initiative, through a grant from the ETH Domain and computational resources provided by the Swiss National Supercomputing Centre (CSCS) under the Alps infrastructure.

Appendix

This appendix provides experimental details, results, and analyses that complement the main paper. • Section A describes the training data, implementation details, simulation and real-world evaluation settings, baseline settings, and inference benchmark protocol.

• Section B presents additional simulation benchmark results on LIBERO, LIBERO-Plus, and RoboCasa. • Section C provides additional ablations and analyses, including backbone variants, pre- training ablations, split-layer analysis, action-token attention, and robustness trends.

Experimental Settings And Reproducibility Details

GAM is trained in two stages, following standard practice for generalist robot policies [2, 4]. The first stage jointly trains the predictor, action head, and the GFM backbone end-to-end on a large mixture of single-arm robot data. The model was trained using 64 NVIDIA GH200 GPUs witch batch size of 1024, which takes approximately ˜96 hours. The second stage fine-tunes the entire model on each benchmark’s official training set before evaluation. The second stage on simulation benchmark was trained using 16 NVIDIA GH200 GPUs with batch size of 160, which takes ˜48 hours.

Pre-Training Details

We pre-train GAM on a weighted mixture of Open-X Embodiment (OXE), MimicGen , and RoboCasa365 . OXE provides broad real-robot coverage across multiple embodiments and manipulation domains, while MimicGen and RoboCasa365 provide simulation demonstrations with clean geometric supervision. The sampling ratios are 72%, 18%, and 10% for OXE, MimicGen, and RoboCasa365, respectively. For future-depth supervision, we use teacher pseudo-depth for OXE and re-rendered simulator depth for MimicGen and RoboCasa365. Figure 6 shows the dataset mixture used for pre-training.

For OXE, we use the subset whose actions can be mapped to our common control interface and exclude datasets that are incompatible with our action space.. For RoboCasa365, we use only the manipulation-task subset. Across all three sources, we keep the original task language provided by the datasets and do not synthesize additional instructions. The language encoder is kept frozen during pre-training.

All datasets are converted to a common observation and action format before training. Images are resized to 224×224, and the model uses two RGB views when available: an external view and a wrist view. Following Cosmos-policy and π0.5 , standard image augmentations such as random cropping, rotation, and color jitter are applied during training and disabled during evaluation.

Simulation Experiments Details

We adopt the LIBERO evaluation protocol established by OpenVLA and OpenVLA-OFT . Specifically, we evaluate on the four standard LIBERO task suites: LIBERO-Spatial, LIBERO- Object, LIBERO-Goal, and LIBERO-Long, each consisting of 10 tasks. Following OpenVLA- OFT, we train on filtered LIBERO demonstrations by removing unsuccessful episodes and filtering idle/no-op frames, i.e., training samples with near-zero actions. We fine-tune a separate policy for each LIBERO suite.

We report task execution success rate (SR, %) as our primary evaluation metric. For the original LIBERO benchmark, each task is evaluated over 50 randomized trials, resulting in 500 rollouts per suite. For LIBERO-Plus, we follow the official evaluation setting and use one rollout per perturbed

0.05

Figure 6: Training Dataset Mixture. We illustrate the dataset mixture utilized during pretraining, detailing the relative proportions of each constituent dataset. The pie chart shows the high-level source mixture, and the bar chart shows the percentage of each constituent dataset relative to the entire training corpus.

Table 5: Hyperparameters for LIBERO, LIBERO-Plus and real-world experiments.

Learning Rate (Lr)

5.16e-5 backbone; 5.16e-4 action head and predictor

Input Images

1 external camera image, 1 wrist-mounted camera image

Action Chunk Size

8 steps (predict 8, execute all 8 open-loop at test time)

Image Augmentations

90% random crop, ±5◦rotation, color jitter, JPEG q95: random resized crop=dict(scale=[0.9, 0.9], ratio=[1.0, 1.0])

Jpeg Quality=95

task instance. All models are trained with a global batch size of 160 for up to 110k training steps, until convergence is achieved. For baseline comparison, we re-evaluate π0.5 and Cosmos-Policy under our evaluation setting using publicly available checkpoints. We also re-evaluate π0.5 + Spatial Forcing and π0.5 + ROCKET using reproduced checkpoints from ROCKET . Results for the remaining baselines are taken from Fei et al. and Zheng et al. .

Related Journal Articles & DOI Links

Selected peer-reviewed publications relevant to 12 Lead ECG Acquisition. Click the DOI to access the full paper (may require institutional access).

Why Choose Us?

Bangalore guidance for robotics, Spectre and autonomous systems projects.