Use HF Papers

lerobot/common/policies/act/modeling_act.py

@@ -15,7 +15,7 @@
 # limitations under the License.
 """Action Chunking Transformer Policy
 
-As per Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware (https://arxiv.org/abs/2304.13705).
+As per Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware (https://huggingface.co/papers/2304.13705).
 The majority of changes here involve removing unused code, unifying naming, and adding helpful comments.
 """
 
@@ -41,7 +41,7 @@
 class ACTPolicy(PreTrainedPolicy):
     """
     Action Chunking Transformer Policy as per Learning Fine-Grained Bimanual Manipulation with Low-Cost
-    Hardware (paper: https://arxiv.org/abs/2304.13705, code: https://github.com/tonyzhaozh/act)
+    Hardware (paper: https://huggingface.co/papers/2304.13705, code: https://github.com/tonyzhaozh/act)
     """
 
     config_class = ACTConfig
@@ -161,7 +161,7 @@ def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict]:
             # Calculate Dₖₗ(latent_pdf || standard_normal). Note: After computing the KL-divergence for
             # each dimension independently, we sum over the latent dimension to get the total
             # KL-divergence per batch element, then take the mean over the batch.
-            # (See App. B of https://arxiv.org/abs/1312.6114 for more details).
+            # (See App. B of https://huggingface.co/papers/1312.6114 for more details).
             mean_kld = (
                 (-0.5 * (1 + log_sigma_x2_hat - mu_hat.pow(2) - (log_sigma_x2_hat).exp())).sum(-1).mean()
             )
@@ -175,7 +175,7 @@ def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict]:
 
 class ACTTemporalEnsembler:
     def __init__(self, temporal_ensemble_coeff: float, chunk_size: int) -> None:
-        """Temporal ensembling as described in Algorithm 2 of https://arxiv.org/abs/2304.13705.
+        """Temporal ensembling as described in Algorithm 2 of https://huggingface.co/papers/2304.13705.
 
         The weights are calculated as wᵢ = exp(-temporal_ensemble_coeff * i) where w₀ is the oldest action.
         They are then normalized to sum to 1 by dividing by Σwᵢ. Here's some intuition around how the

lerobot/common/policies/diffusion/configuration_diffusion.py

@@ -81,7 +81,7 @@ class DiffusionConfig(PreTrainedConfig):
         n_groups: Number of groups used in the group norm of the Unet's convolutional blocks.
         diffusion_step_embed_dim: The Unet is conditioned on the diffusion timestep via a small non-linear
             network. This is the output dimension of that network, i.e., the embedding dimension.
-        use_film_scale_modulation: FiLM (https://arxiv.org/abs/1709.07871) is used for the Unet conditioning.
+        use_film_scale_modulation: FiLM (https://huggingface.co/papers/1709.07871) is used for the Unet conditioning.
             Bias modulation is used be default, while this parameter indicates whether to also use scale
             modulation.
         noise_scheduler_type: Name of the noise scheduler to use. Supported options: ["DDPM", "DDIM"].

lerobot/common/policies/diffusion/modeling_diffusion.py

@@ -48,7 +48,7 @@
 class DiffusionPolicy(PreTrainedPolicy):
     """
     Diffusion Policy as per "Diffusion Policy: Visuomotor Policy Learning via Action Diffusion"
-    (paper: https://arxiv.org/abs/2303.04137, code: https://github.com/real-stanford/diffusion_policy).
+    (paper: https://huggingface.co/papers/2303.04137, code: https://github.com/real-stanford/diffusion_policy).
     """
 
     config_class = DiffusionConfig
@@ -370,7 +370,7 @@ def compute_loss(self, batch: dict[str, Tensor]) -> Tensor:
 class SpatialSoftmax(nn.Module):
     """
     Spatial Soft Argmax operation described in "Deep Spatial Autoencoders for Visuomotor Learning" by Finn et al.
-    (https://arxiv.org/pdf/1509.06113). A minimal port of the robomimic implementation.
+    (https://huggingface.co/papers/1509.06113). A minimal port of the robomimic implementation.
 
     At a high level, this takes 2D feature maps (from a convnet/ViT) and returns the "center of mass"
     of activations of each channel, i.e., keypoints in the image space for the policy to focus on.
@@ -728,7 +728,7 @@ def __init__(
 
         self.conv1 = DiffusionConv1dBlock(in_channels, out_channels, kernel_size, n_groups=n_groups)
 
-        # FiLM modulation (https://arxiv.org/abs/1709.07871) outputs per-channel bias and (maybe) scale.
+        # FiLM modulation (https://huggingface.co/papers/1709.07871) outputs per-channel bias and (maybe) scale.
         cond_channels = out_channels * 2 if use_film_scale_modulation else out_channels
         self.cond_encoder = nn.Sequential(nn.Mish(), nn.Linear(cond_dim, cond_channels))
 

lerobot/common/policies/pi0fast/modeling_pi0fast.py

@@ -17,7 +17,7 @@
 """
 π0+FAST: Efficient Action Tokenization for Vision-Language-Action Models
 
-[Paper](https://arxiv.org/abs/2501.09747)
+[Paper](https://huggingface.co/papers/2501.09747)
 [Jax code](https://github.com/Physical-Intelligence/openpi)
 
 Designed by Physical Intelligence. Ported from Jax by Hugging Face.

lerobot/common/policies/tdmpc/modeling_tdmpc.py

@@ -17,8 +17,8 @@
 """Implementation of Finetuning Offline World Models in the Real World.
 
 The comments in this code may sometimes refer to these references:
-    TD-MPC paper: Temporal Difference Learning for Model Predictive Control (https://arxiv.org/abs/2203.04955)
-    FOWM paper: Finetuning Offline World Models in the Real World (https://arxiv.org/abs/2310.16029)
+    TD-MPC paper: Temporal Difference Learning for Model Predictive Control (https://huggingface.co/papers/2203.04955)
+    FOWM paper: Finetuning Offline World Models in the Real World (https://huggingface.co/papers/2310.16029)
 """
 
 # ruff: noqa: N806

lerobot/common/policies/vqbet/modeling_vqbet.py

@@ -162,7 +162,7 @@ def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict]:
         batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
         batch["observation.images"] = torch.stack([batch[key] for key in self.config.image_features], dim=-4)
         batch = self.normalize_targets(batch)
-        # VQ-BeT discretizes action using VQ-VAE before training BeT (please refer to section 3.2 in the VQ-BeT paper https://arxiv.org/pdf/2403.03181)
+        # VQ-BeT discretizes action using VQ-VAE before training BeT (please refer to section 3.2 in the VQ-BeT paper https://huggingface.co/papers/2403.03181)
         if not self.vqbet.action_head.vqvae_model.discretized.item():
             # loss: total loss of training RVQ
             # n_different_codes: how many of the total possible VQ codes are being used in single batch (how many of them have at least one encoder embedding as a nearest neighbor). This can be at most `vqvae_n_embed * number of layers of RVQ (=2)`.
@@ -185,7 +185,7 @@ def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict]:
 class SpatialSoftmax(nn.Module):
     """
     Spatial Soft Argmax operation described in "Deep Spatial Autoencoders for Visuomotor Learning" by Finn et al.
-    (https://arxiv.org/pdf/1509.06113). A minimal port of the robomimic implementation.
+    (https://huggingface.co/papers/1509.06113). A minimal port of the robomimic implementation.
 
     At a high level, this takes 2D feature maps (from a convnet/ViT) and returns the "center of mass"
     of activations of each channel, i.e., keypoints in the image space for the policy to focus on.
@@ -387,7 +387,7 @@ def forward(self, batch: dict[str, Tensor], rollout: bool) -> tuple[dict, dict]:
 
         # only extract the output tokens at the position of action query:
         # Behavior Transformer (BeT), and VQ-BeT are both sequence-to-sequence prediction models,
-        # mapping sequential observation to sequential action (please refer to section 2.2 in BeT paper https://arxiv.org/pdf/2206.11251).
+        # mapping sequential observation to sequential action (please refer to section 2.2 in BeT paper https://huggingface.co/papers/2206.11251).
         # Thus, it predicts a historical action sequence, in addition to current and future actions (predicting future actions : optional).
         if len_additional_action_token > 0:
             features = torch.cat(
@@ -824,8 +824,8 @@ def get_action_from_latent(self, latent):
             return einops.rearrange(output, "N (T A) -> N T A", A=self.config.action_feature.shape[0])
 
     def get_code(self, state):
-        # in phase 2 of VQ-BeT training, we need a `ground truth labels of action data` to calculate the Focal loss for code prediction head. (please refer to section 3.3 in the paper https://arxiv.org/pdf/2403.03181)
-        # this function outputs the `GT code` of given action using frozen encoder and quantization layers. (please refer to Figure 2. in the paper https://arxiv.org/pdf/2403.03181)
+        # in phase 2 of VQ-BeT training, we need a `ground truth labels of action data` to calculate the Focal loss for code prediction head. (please refer to section 3.3 in the paper https://huggingface.co/papers/2403.03181)
+        # this function outputs the `GT code` of given action using frozen encoder and quantization layers. (please refer to Figure 2. in the paper https://huggingface.co/papers/2403.03181)
         state = einops.rearrange(state, "N T A -> N (T A)")
         with torch.no_grad():
             state_rep = self.encoder(state)
@@ -838,7 +838,7 @@ def get_code(self, state):
             return state_vq, vq_code
 
     def vqvae_forward(self, state):
-        # This function passes the given data through Residual VQ with Encoder and Decoder. Please refer to section 3.2 in the paper https://arxiv.org/pdf/2403.03181).
+        # This function passes the given data through Residual VQ with Encoder and Decoder. Please refer to section 3.2 in the paper https://huggingface.co/papers/2403.03181).
         state = einops.rearrange(state, "N T A -> N (T A)")
         # We start with passing action (or action chunk) at:t+n through the encoder ϕ.
         state_rep = self.encoder(state)

lerobot/common/policies/vqbet/vqbet_utils.py

@@ -336,7 +336,7 @@ class ResidualVQ(nn.Module):
     """
     Residual VQ is composed of multiple VectorQuantize layers.
 
-    Follows Algorithm 1. in https://arxiv.org/pdf/2107.03312.pdf
+    Follows Algorithm 1. in https://huggingface.co/papers/2107.03312
         "Residual Vector Quantizer (a.k.a. multi-stage vector quantizer [36]) cascades Nq layers of VQ as follows. The unquantized input vector is
         passed through a first VQ and quantization residuals are computed. The residuals are then iteratively quantized by a sequence of additional
         Nq -1 vector quantizers, as described in Algorithm 1."
@@ -1006,7 +1006,7 @@ def gumbel_sample(
     if not straight_through or temperature <= 0.0 or not training:
         return ind, one_hot
 
-    # use reinmax for better second-order accuracy - https://arxiv.org/abs/2304.08612
+    # use reinmax for better second-order accuracy - https://huggingface.co/papers/2304.08612
     # algorithm 2
 
     if reinmax:
@@ -1156,7 +1156,7 @@ def batched_embedding(indices, embeds):
 
 
 def orthogonal_loss_fn(t):
-    # eq (2) from https://arxiv.org/abs/2112.00384
+    # eq (2) from https://huggingface.co/papers/2112.00384
     h, n = t.shape[:2]
     normed_codes = F.normalize(t, p=2, dim=-1)
     cosine_sim = einsum("h i d, h j d -> h i j", normed_codes, normed_codes)