Open Models Almanac
ModelsXiaomi-Robotics-1

Xiaomi

Xiaomi-Robotics-1

What Xiaomi-Robotics-1’s own documents say it is built from — every method with the sentence that describes it.

Attention
Qwen3-VL backbone
Input
text+image -> action
Licence
Apache-2.0
Released
2026-07-28

Against the consensus recipe

Its documents state 0 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

Checked in its code

Its code shows 3 of the 19 architecture features checked — 3 of them are not stated in its documents.

Sources say what a lab chose to describe; the modeling code says what the checkpoint runs. Each feature below was put to a code verifier, and only the config values and source lines it cited that were found in the files are shown, linked to the line.

Read from the model repository’s own code (modeling_mibot.py, configuration_mibot.py) and config.json at revision 3a6d0293bf, by anthropic/claude-sonnet-5 on 2026-11-07.

FeatureCodeIts documentsVerdict
Grouped-query attention code only in its codenot stated

The Qwen3VL text attention uses 32 query heads with only 8 key/value heads, repeated via repeat_kv for grouped-query attention.

show evidencevlm_config.text_config.num_attention_heads = 32vlm_config.text_config.num_key_value_heads = 8
self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
modeling_mibot.py · L406
Multi-head latent attention not in its codenot stated

No low-rank KV latent projection (kv_lora_rank/q_lora_rank) exists anywhere in the attention code or config; attention uses standard…

show the whole verdict

No low-rank KV latent projection (kv_lora_rank/q_lora_rank) exists anywhere in the attention code or config; attention uses standard per-head q/k/v projections.

Sliding-window attention not in its codenot stated

No sliding_window config key or windowed-mask code is present for the language model; attention is full causal attention.

Interleaved sliding-window and global attention not in its codenot stated

There is no layer_types or sliding_window_pattern config, and no code that interleaves local and global attention layers.

Indexer-selected sparse attention (DeepSeek Sparse Attention) not in its codenot stated

No indexer scoring mechanism or index_topk/index_n_heads/index_head_dim config exists in this model.

Linear-attention or state-space layers alongside full attention not in its codenot stated

All decoder layers use the same Qwen3VLTextAttention softmax attention; no linear-attention or SSM layer variants exist.

Gated DeltaNet layers not in its codenot stated

No gated delta-rule recurrent mixer or linear_num_key_heads/linear_conv_kernel_dim config is present in the code.

Mamba-2 layers not in its codenot stated

No Mamba-2 SSD blocks or ssm_state_size/mamba_num_heads config exist; the model is a standard transformer decoder.

Learnable attention sink not in its codenot stated

The DiT policy head has a learned 'sink' embedding concatenated as a token, not a per-head sink logit added to the attention softmax…

show the whole verdict

The DiT policy head has a learned 'sink' embedding concatenated as a token, not a per-head sink logit added to the attention softmax denominator; the language model's attention has no sink logit mechanism at all.

Gated attention output not in its codenot stated

Neither the Qwen3VL text attention nor the DiT attention multiplies the attention output by an input-dependent sigmoid gate before the…

show the whole verdict

Neither the Qwen3VL text attention nor the DiT attention multiplies the attention output by an input-dependent sigmoid gate before the output projection; AdaLN gates (gate_msa/gate_mlp) scale the residual branch output, not an attention-internal gate, and apply to the action DiT head, not the language model.

QK normalization code only in its codenot stated

Qwen3VLTextAttention applies RMSNorm to queries and keys per head before rotary embedding and the attention dot product.

show evidence
self.q_norm = Qwen3VLTextRMSNorm(self.head_dim, eps=config.rms_norm_eps)  # unlike olmo, only on the head dim!
        self.k_norm = Qwen3VLTextRMSNorm(
            self.head_dim, eps=config.rms_norm_eps
        )  # thus post q_norm does not need reshape
modeling_mibot.py · L423–426
query_states = self.q_norm(self.q_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
        key_states = self.k_norm(self.k_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
modeling_mibot.py · L440–441
Partial RoPE not in its codenot stated

RoPE is applied to the full head dimension via rotate_half over the whole head_dim with no partial_rotary_factor or split dimension; no…

show evidence

RoPE is applied to the full head dimension via rotate_half over the whole head_dim with no partial_rotary_factor or split dimension; no mechanism reduces the rotated portion.

def rotate_half(x):
    """Rotates half the hidden dims of the input."""
    x1 = x[..., : x.shape[-1] // 2]
    x2 = x[..., x.shape[-1] // 2 :]
    return torch.cat((-x2, x1), dim=-1)
modeling_mibot.py · L122–126
YaRN RoPE scaling not in its codenot stated

rope_scaling in text_config sets rope_type to 'default' with mrope_section/mrope_interleaved for multimodal RoPE, not YaRN, so YaRN…

show evidence

rope_scaling in text_config sets rope_type to 'default' with mrope_section/mrope_interleaved for multimodal RoPE, not YaRN, so YaRN scaling is not enabled.

vlm_config.text_config.rope_scaling = {"mrope_interleaved": true, "mrope_section": [24, 20, 20], "rope_type": "defa…
Layers without positional encoding (NoPE) not in its codenot stated

Every Qwen3VLTextDecoderLayer uses the same attention class with RoPE applied uniformly; there is no per-layer switch omitting positional…

show the whole verdict

Every Qwen3VLTextDecoderLayer uses the same attention class with RoPE applied uniformly; there is no per-layer switch omitting positional encoding.

Mixture of experts not in its codenot stated

No router, num_experts, or expert MLP module exists; Qwen3VLTextMLP is a single dense SwiGLU MLP per layer.

Shared expert not in its codenot stated

There is no MoE structure at all, so no shared-expert mechanism or related config keys exist.

Auxiliary-loss-free load balancing (selection bias) not in its codenot stated

No router or e_score_correction_bias exists since the model has no mixture-of-experts feed-forward layers.

Multi-token prediction layers not in its codenot stated

No num_nextn_predict_layers or MTP module declarations exist in the config or code; the model only outputs a single next-token/action…

show the whole verdict

No num_nextn_predict_layers or MTP module declarations exist in the config or code; the model only outputs a single next-token/action prediction head.

SwiGLU feed-forward code only in its codenot stated

Qwen3VLTextMLP computes silu(gate_proj(x)) * up_proj(x) then down_proj, a SwiGLU feed-forward, matching hidden_act 'silu' in the text…

show evidence

Qwen3VLTextMLP computes silu(gate_proj(x)) * up_proj(x) then down_proj, a SwiGLU feed-forward, matching hidden_act 'silu' in the text config.

vlm_config.text_config.hidden_act = "silu"
def forward(self, x):
        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
        return down_proj
modeling_mibot.py · L483–485

filed at the root 1

Mixture-of-Transformers (MoT) core

“Xiaomi-Robotics-1 adopts a Mixture-of-Transformers (MoT) architecture consisting of a pre-trained vision-language model (VLM) (i.e.…”

token mixer 3

softmax attention 3

Causal attention used

“To account for the temporal relationship between actions at different timesteps, we use causal attention in the DiT.”

Shape attention mask for action-prefix conditioning used

“we change the original causal attention mask of the DiT to a -shape attention mask [20, 71, 16]”

attention sink 1

Learnable attention sink token used

“We add a learnable attention sink token at the front of the state and noisy action tokens to stabilize the attention distribution during…”

positional encoding 1

RoPE positional-index offset used

“We first simply add an offset to the RoPE positional indices of the noisy action tokens to enable the model to distinguish tokens of noisy…”

normalization & residual 1

Adaptive Layer Normalization (AdaLN) used

“We leverage adaptive normalization layers (adaLN) [49, 48] to inject the flow-matching timestep condition into the DiT for action…”

multimodal architecture 1

Masking action-related tokens from DiT attention used

“we exclude these action-related tokens from the DiT’s attention computation, constraining the DiT tokens to attend solely to the…”

filed at the root 1

Winner-takes-all action-candidate supervision used

“Action prediction is supervised via a winner-takes-all scheme: only the candidate with the lowest distance is updated via backpropagation.”

language modelling objective 1

Chain-of-thought labeling used

“We leverage chain-of-thought (CoT) labeling as in ERVLA [61] and train our model with a 50% probability on the next-token-prediction loss…”

auxiliary loss 3

Beta-distributed flow-matching timestep sampling used

“we sample from a Beta distribution, placing more weight on noisier timesteps during training.”

Auxiliary action-generation supervision for the vision-language model used

“we introduce an auxiliary action-generation supervision on the VLM.”

Loss masking for missing embodiment-specific action dimensions used

“We mask out the dimensions corresponding to missing action components during loss computation.”

training runtime 1

Asynchronous fine-tuning used

“We fine-tune Xiaomi-Robotics-1 on the data of all four tasks using the asynchronous training recipe proposed in [8].”

data filtering 1

Filtering idle trajectory segments used

“We filter out idle segments within trajectories to prevent the model from learning uninformative or noisy signals.”

synthetic data 5

Automated language annotation of manipulation trajectories core

“we develop an auto-labeling pipeline that first divides each trajectory into equal-length segments and leverage Qwen3.5-27B [70] to…”

Cross-validated consensus annotation for visual grounding used

“For visual grounding, we develop a rigorous cross-validated consensus mechanism that integrates Grounded SAM [55], Grounding DINO 1.5…”

Language annotation of human-segmented post-training clips used

“We leverage Qwen3.5 [70] to annotate human-segmented video clips with language instructions.”

Pre-trained VLM relabeling to refine VQA and caption annotations used

“VQA and captioning annotation quality is further refined through re-labeling using state-of-the-art pre-trained VLMs [1].”

VLM-generated embodied reasoning and planning data from robot trajectories used

“For embodied reasoning & planning, we leverage pre-trained VLMs to generate data from root trajectories”

data mixture & curriculum 3

1:6 vision-language-to-robot-trajectory sampling used

“We sample vision-language data and robot trajectory data at a ratio of 1:6.”

Joint training on robot trajectories and vision-language data used

“we co-train the model with the entire robot trajectory data and vision-language data”

Joint vision-language and action training used

“we further co-train the model on a high-quality vision-language dataset curated in our previous work [8] under the next-token prediction…”

sequence packing 1

Vision-language token sequence packing used

“we pack all vision-language tokens within a batch into a single sequence for a VLM forward pass.”

filed at the root 1

Pre-training followed by post-training for robot policies core

“We adopt a two-stage training recipe consisting of pre-training and post-training.”

inference scheduling 2

Asynchronous inference and execution core

“Asynchronous execution enables the robot to continue rolling out trajectories during model inference.”

Action-chunk timestep alignment used

“During deployment, we carefully align the timesteps of action chunks generated from consecutive inferences to ensure continuous and…”

inference kernel 1

FlashAttention-2 used

“attn_implementation="flash_attention_2",”

context management 1

Four-step observation history used

“observation history: 4”

unfiled 18

Flow matching core

“the DiT generates the action chunk via flow-matching [49]”

Joint flow-matching and VLM choice-policy training objective core

“The model is optimized to predict actions by jointly minimizing the flow-matching loss of the DiT and the regression loss of the VLM…”

Choice Policies for multimodal action prediction used

“To account for the multi-modality in trajectories, we adopt Choice Policies [51] for action prediction.”

Choice Policies for VLM action generation used

“we leverage Choice Policies [59] to enable action generation directly within the VLM framework [8].”

Cross-embodiment end-effector frame orientation alignment used

“we unify the orientation of the end-effector frames across all robot data and UMI data in both the pre-training and post-training datasets.”

Eight-server launcher for reproducible multi-GPU evaluation used

“For reproducible multi-GPU evaluation, use the standard eight-server launcher in the companion source repository.”

Five-step Euler integration for action sampling used

“The clean action chunk is recovered via a 5-step Euler integration”

Fixed-length segmentation of manipulation trajectories used

“first divides each trajectory into equal-length segments”

Four-timestep sampling to amortize VLM computation used

“we amortize its cost by sampling four flow-matching timesteps per sample.”

Load custom Transformers code with trust_remote_code enabled used

“The custom model and processor must be loaded with trust_remote_code=True.”

Online-error-weighted flow-matching loss used

“we dynamically re-weight the flow-matching loss based on the error between the online-predicted actions and the ground-truth actions.”

Predict sixteen actions per query used

“actions per query: 16”

Prefixing the next inference with the current chunk's actions across the inference window used

“We condition the next inference by prefixing the noisy actions with the actions from step to step of the current chunk.”

Producer-consumer parallel pipeline for trajectory captioning used

“we develop a producer–consumer pipeline that decouples clip segmentations from caption labeling: while CPU worker threads cut per-segment…”

Relative-delta end-effector pose action representation used

“For arm actions, we adopt relative delta end-effector (EE) poses with respect to the current state”

Set the observation interval to two used

“observation interval: 2”

Timestamp-based 30 Hz synchronization of input modalities used

“we synchronize all input modalities by resampling them to a unified 30Hz timeline using timestamps.”

Training-time action-prefix conditioning for asynchronous execution used

“we follow training RTC [5] and condition action generation on previously committed actions by prefixing them to the noisy action tokens in…”

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/XiaomiRobotics/Xiaomi-Robotics-1-RoboCasa365model cardXiaomiread
arxiv.org/abs/2607.15330technical reportXiaomiread
arxiv.org/abs/2602.12684technical reportXiaomiread
↑