Open Models Almanac
ModelsAlpamayo 1.5 / 2 Super

NVIDIA

Alpamayo 1.5 / 2 Super

What Alpamayo 1.5 / 2 Super’s own documents say it is built from — every method with the sentence that describes it.

Parameters
10B
Attention
Cosmos-Reason2-8B backbone; reasoning text + driving trajectory
Input
text+image -> action
Licence
OpenMDW-1.1
Released
2026-03-03

Against the consensus recipe

Its documents state 1 of the 18 methods the field agrees on.

The recipe from the overview: the methods the most labs adopt, per stage. A missing one is something its documents do not say, not something the model lacks.
  1. Data curation

  2. Training objective

  3. Model architecture

  4. Optimization

  5. Post-training

  6. Inference & serving

prediction head 1

Diffusion-based trajectory decoder core

“featuring a diffusion-based trajectory decoder”

multimodal architecture 3

Separate flow-matching action expert core

“we adopt a separate action-expert to decode actions via flow matching”

Chain-of-thought reasoning vision-language-action model core

“NVIDIA’s open 10B-parameter chain-of-thought reasoning VLA model”

End-to-end autonomous driving used

“E2E approaches map raw sensor inputs directly to vehicle motion through jointly trained neural networks.”

1 more mentioned or evaluated
Triplane-based multi-camera tokenization evaluated

multimodal architecture “the triplane-based multi-camera tokenizer from Ivanovic et al. (2025) achieves nearly identical minADE6 values as the baseline”

language modelling objective 1

Joint cross-entropy training of reasoning and trajectory tokens used

“we apply the cross-entropy loss over both the reasoning tokens and the discrete trajectory tokens”

auxiliary loss 2

Knowledge insulation used

“During training, we apply a stop-gradient to the KV-cache produced by the VLM to prevent gradients from the expert back-propagating into…”

KL regularization to a reference policy used

“The KL regularization term with coefficient penalizes deviations from the reference policy (typically the SFT model)”

1 more mentioned or evaluated
Autoregressive discrete-action prediction evaluated

language modelling objective “a baseline model trained to auto-regressively predict 6 discrete trajectory tokens”

filed at the root 1

Chain of Causation (CoC) labeling framework core

“We develop a structured Chain of Causation (CoC) labeling framework that produces decision-grounded, causally-linked reasoning traces”

data filtering 7

Disagreement-based curation of RL post-training data used

“We therefore prioritize these high-disagreement samples to construct a focused post-training dataset, while mixing in a similar proportion…”

Independent annotator quality checks used

“Each labeled instance first undergoes a quality check performed by a different annotator.”

Keyframe selection for causal reasoning annotation used

“Selecting keyframes for CoC annotation is therefore critical to maximizing the clarity of decision causality.”

Keyframe-range annotation for proactive scenarios used

“For proactive scenarios, we annotate a keyframe range: a time window during which the ego actively evaluates or prepares for a potential…”

Rule-based meta-action detection used

“implement corresponding rule-based detectors to infer these meta actions at the frame level.”

Rule-based scenario clip selection and balancing used

“We employ rule-based methods to identify clips corresponding to each scenario and balance the number of clips per scenario to ensure…”

Safety exclusion filtering of invalid annotations used

“apply a safety exclusion filter to remove invalid data with illegal or unsafe driving behavior”

synthetic data 4

Causally grounded, decision-linked reasoning traces core

“We first define a comprehensive set of high-level driving decisions that directly correspond to low-level ego trajectories. Each reasoning…”

Temporal separation of historical and future inputs in CoC auto-labeling used

“VLMs are prompted to use the 2-second historical video when identifying critical components.”

VLM-based automatic labeling with heuristic rules used

“Automated (VLM-based auto-labeling and heuristic rules)”

VLM-based offline automatic CoC labeling used

“we employ state-of-the-art VLMs such as GPT-5 (OpenAI, 2025) to perform offline auto-labeling through a multi-step reasoning process.”

tokenization 2

Single-image tokenization for multi-camera video inputs default

“AR1 adopts single-image tokenization by default”

Discrete action-token representation used

“During training, we inject the action modality to the VLM through discrete tokens”

filed at the root 2

Multi-stage training strategy core

“We adopt a multi-stage training strategy that builds upon the Cosmos-Reason VLM backbone”

Three-stage VLA training: action-modality injection, CoC reasoning SFT, and RL post-training core

“we adopt a three-stage training strategy to transform the VLM into a reasoning-capable autonomous driving policy.”

supervised fine-tuning 1

Supervised fine-tuning on Chain of Causation data core

“We perform SFT on the CoC dataset to teach the model to generate reasoning traces through imitation”

reinforcement learning algorithm 3

RL-based post-training framework with complementary rewards core

“we introduce an RL-based post-training framework shown in Fig. 6 that optimizes three complementary reward signals”

Reinforcement learning used

“we apply RL-based post-training to simultaneously improve reasoning quality, reasoning-action consistency, and trajectory quality”

Group Relative Policy Optimization used

“We adopt GRPO (Shao et al., 2024) as our alignment algorithm.”

reward modelling 4

Reasoning reward from LRM feedback used

“we introduce a reasoning reward derived from LRM feedback”

Reasoning-action consistency reward used

“we introduce a CoC–action consistency reward that measures behavioral alignment between the generated reasoning trace and the…”

Safety reward used

“we introduce a safety reward that penalizes unsafe or physically implausible trajectories during post-training”

Trajectory-quality reward used

“The reward combines three terms:”

mid-training & continual pretraining 1

Physical-AI-focused pretraining used

“Alpamayo-R1 adopts Cosmos-Reason as its VLM backbone, specifically post-trained on Physical AI data including driving scenarios.”

filed at the root 1

Structured True/False LLM evaluation used

“By reformulating these aspects as a set of True/False questions, the evaluation process becomes more interpretable and better aligned with…”

judge 1

Large reasoning model as evaluator used

“we employ large reasoning models (LRMs) as automatic evaluators to provide scalable, high-quality feedback on reasoning quality.”

human & real-world evaluation 3

Closed-loop simulation used

“Closed-Loop Evaluation using AlpaSim on 913 scenarios”

Human-verified and LLM-based evaluation used

“we adopt a hybrid evaluation strategy that combines human verification with LLM-based auto-evaluation.”

On-vehicle road testing used

“Proprietary autonomous driving test datasets, closed-loop simulation, on-vehicle road tests.”

filed at the root 3

Decision-grounded keyframe selection used

“For reactive scenarios, a keyframe is typically chosen by applying a short temporal buffer (approximately 0.5 seconds) before the ego…”

Two-stage human CoC annotation used

“human annotators are required to complete a two-stage procedure designed to produce concise and causally grounded CoC write-ups.”

Unicycle-dynamics control representation used

“we adopt an action representation governed by unicycle dynamics that leads to better closed-loop performance”

unfiled 19

Chain-of-Causation reasoning core

“Both models generate Chain-of-Causation reasoning traces alongside trajectory predictions.”

Flow-matching diffusion action expert for trajectory decoding core

“We employ a diffusion-based action-expert trajectory decoder built on flow matching”

Hybrid human and automatic labeling core

“we develop a hybrid labeling procedure that combines human labeling and auto-labeling.”

Hybrid reasoning-trace labeling with automated labeling and human review core

“constructed through a hybrid labeling pipeline that combines large-scale auto-labeling with humans in the loop”

Flow matching used

“During inference, we set , i.e., 5 steps, in flow matching to reduce latency with negligible performance degradation.”

Closed-set decision labels aligned with ego trajectories used

“we define a closed set of high-level driving decisions as in Tab. 1. Each clip is annotated with at most one longitudinal and one lateral…”

Conditional flow matching objective for the action expert used

“We train the action-expert using a vanilla conditional flow matching loss”

Continuous trajectory decoding with unicycle dynamics and flow matching used

“adopting a continuous action representation governed by unicycle dynamics with flow matching”

Euler integration for flow-matching trajectory denoising used

“During inference, starting with , we perform denoising through Euler integration:”

Gaussian conditional optimal-transport path for flow matching used

“In practice, we adopt the Gaussian conditional optimal transport (OT) path”

Hybrid human-in-the-loop and automated labeling used

“supported by a hybrid human-in-the-loop and auto-labeling pipeline for scalable high-quality data generation”

Joint discrete trajectory-token and continuous flow-matching action-expert representation used

“combining discrete trajectory tokens learned within the VLM with an action-expert that decodes the same trajectories into continuous…”

Pre-action keyframe placement using a short temporal buffer used

“a keyframe is typically chosen by applying a short temporal buffer (approximately 0.5 seconds) before the ego vehicle initiates a behavior…”

Rule-based matching of inferred intent against trajectory-derived meta-actions used

“compare it against the meta-actions derived from the predicted trajectory using rule-based matching.”

Tikhonov-regularized least-squares derivation of control sequences used

“the ground-truth control sequence is derived from through a least-squares formulation with Tikhonov regularization to attenuate…”

4 more mentioned or evaluated
Flex multi-camera vision tokenization evaluated

“Flex (Yang et al., 2025) is able to achieve more drastic improvements”

Scaling vision-language model backbone capacity evaluated

“we observe consistent improvements in open-loop performance as model size increases”

Training data scaling evaluated

“We train the 0.5B model on varying amounts of data: 100k, 200k, 500k, 1M, and 2M video segments, keeping the total number of training…”

V-model iterative unit- and system-level testing and validation mentioned

“Following the V-model methodology, iterative testing and validation at both unit and system levels are essential”

Sources

The curated source list for this model. "not fetched" means the URL is recorded but its text was not read in the current run.

SourceKindPublisherStatus
huggingface.co/nvidia/Alpamayo-1.5-10Bmodel cardNVIDIAread
arxiv.org/abs/2511.00088technical reportNVIDIAread
↑