2.1. RL formulation for learning image-based robotic manipulation tasks

Agent and environment interaction is formally defined by the Markov Decision Process (MDP). A Markov Decision Process is a discrete-time stochastic control process. In RL, we formally define the MDP as a tuple $\langle \mathcal{S},\mathcal{A}, \mathcal{P}, \mathcal{R}, \gamma \rangle$ . Here, $\mathcal{S}$ is a finite set of states, $\mathcal{A}$ is a finite set of actions, $\mathcal{P}$ is the state transition probability matrix, $\mathcal{R}$ set of rewards for all state-action pair and $\gamma$ is the discount factor. A stochastic policy is defined as a probability distribution over actions, given the states, that is, the probability of taking each action for every state. $\pi (\boldsymbol{a}_t|\boldsymbol{s}_t) = \mathbb{P}[\boldsymbol{a}_t \in \mathcal{A}|\boldsymbol{s}_t \in \mathcal{S}]$ . We formulate the vision-based robotic manipulation tasks in the MDP framework as below.

• Environment: It consists of WidowX 250, a five-axe robot arm equipped with a gripper. We place a table in front of the robot and a camera in the environment in an eye-to-hand configuration. Every task consists of an object placed on the table, which needs to be manipulated to complete the task successfully.

• State: The state $\boldsymbol{s}_t$ represents the RGB image of the environment captured at time step $t$ . We use images of size $48\times 48\times 3$ .

• Action: We define the action at the time step $t$ as a 7-dimensional vector $\boldsymbol{a}_t=\begin{bmatrix}\Delta \boldsymbol{x}_t&\Delta \boldsymbol{o}_t&g_t\end{bmatrix}^\top$ . Here, $\Delta \boldsymbol{x}_t\in \mathbb{R}^3$ , $\Delta \boldsymbol{o}_t\in \mathbb{R}^3$ , $g_t\in \{0,1\}$ denotes the change in position, change in orientation, and gripper command (open/close), respectively, at time step $t$ .

• Reward: The reward $r(\boldsymbol{s}_t,\boldsymbol{a}_t)\in \{0,1\}$ is a binary variable which is equal to 1 if the task is successful and 0, otherwise.

The reward is kept simple and not shaped according to the tasks so that the same reward framework can be used while scaling for a large number of tasks. Also, giving a reward at each time step, instead of at the end of the episode, makes the sum of rewards during an episode dependent on time steps. Therefore, if the agent completes a task in fewer steps, the total reward for that episode will be more.

2.2. Sequential learning problem and solution

We define the sequential tasks learning problem as follows. The agent is required to learn $N$ number of tasks but with the condition that tasks will be given sequentially to the agent and not simultaneously. Therefore, when the agent is learning to perform a particular task, it can only access the data of the current task. This learning process reassembles how a human learns. Let a sequence of robotic manipulation tasks $T_1, T_2, \ldots, T_N$ be given. We assume that each task has the same type of state and action space. Each task has its own data in typical offline reinforcement learning format $\langle \boldsymbol{s}_t, \boldsymbol{a}_t, r_t, \boldsymbol{s}_{t+1}\rangle$ . The agent has to learn a single policy $\pi$ , a mapping from state to action, for all tasks. If we naively train a neural network in a sequential manner, the problem of catastrophic forgetting will occur, which means performance on the previous task will decrease drastically as soon as the neural network starts learning a new task.

We use a regularisation-based approach presented in ref. [Reference Zenke, Poole and Ganguli35] to mitigate the problem of catastrophic forgetting. Fig. 1 describes the framework we developed to solve this problem.

3.1. SAC-CQL algorithm for offline RL

There are two frameworks, namely online and offline learning, to train an RL agent. In the case of an online RL training framework, an RL agent interacts with the environment to collect experience, update itself (train), interact again, and so on. Simply put, the environment is always available for the RL agent to evaluate and improve itself further. This interaction loop is repeated for many episodes during training until the RL agent gets good enough to perform the task successfully. This dynamic approach allows the agent to adapt to unforeseen circumstances but may be computationally expensive and less sample-efficient. While in offline RL settings, we collect data once and then it is not required to interact with the environment. These data can be collected by executing a hand-designed policy or by a human controlling the robot (human demonstration). Data are a sequence of $\langle \boldsymbol{s}_t, \boldsymbol{a}_t, r_t, \boldsymbol{s}_{t+1}\rangle$ tuples. Offline RL poses unique challenges such as data distribution shifts and the need to balance exploration and exploitation without the ability to collect additional data during the learning process. While it may lack the adaptability of online RL, offline RL is computationally efficient and allows for systematically exploring individual tasks with carefully curated datasets.

In recent years, SAC [Reference Haarnoja, Zhou, Abbeel and Levine33] has emerged as one of the robust ways for training RL agents in continuous action space (when action is a real vector), which typically is the case in robotics. SAC is an off-policy entropy-based actor-critic method for continuous action MDPs. Entropy-based methods add entropy term to the existing optimisation goal of maximising expected reward. In addition to maximising expected reward, the RL agent also needs to maximise the entropy of the overall policy. This helps make the policy inherently exploratory and not stuck inside a local minima. Haarnoja et al. [Reference Haarnoja, Zhou, Abbeel and Levine33] define the RL objective in maximum entropy RL settings as in (1).

(1) \begin{equation} J(\pi ) = \sum \limits _{t=0}^{T} \mathbb{E}_{(\mathbf{s}_t, \mathbf{a}_t)\sim \rho _\pi }[r(\mathbf{s}_t, \mathbf{a}_t)+\alpha \mathcal{H}(\pi (\!\cdot |\mathbf{s}_t))]. \end{equation}

Here, $\rho _\pi (\mathbf{s}_t, \mathbf{a}_t)$ denotes the joint distribution of the state and actions over all trajectories that the agent could take and $\mathcal{H}(\pi (\!\cdot |\mathbf{s}_t))$ is the entropy of the policy for state $\mathbf{s}_t$ as defined in (2). $\alpha$ is the temperature parameter controlling the entropy in the policy. $\mathbb{E}$ represents the expectation over all state and action pairs sampled from the trajectory distribution $\rho _\pi$ . Overall, this objective function tries to maximise the expected sum of rewards along with the entropy of the policy.

(2) \begin{equation} \mathcal{H}(\pi (\!\cdot |\mathbf{s}_t)) = \mathbb{E}[-\text{log}(f_\pi (\!\cdot |\mathbf{s}_t))]. \end{equation}

Here, $\pi (\!\cdot |\mathbf{s}_t)$ is a probability distribution over actions and $f_\pi (\!\cdot |\mathbf{s}_t)$ is the probability density function of the policy $\pi$ , we have selected Gaussian distribution to represent the policy.

SAC provides an actor-critic framework where the actor separately represents the policy, and the critic only helps in improving the actor, thus limiting the role of critic only to training. As our state ( $\boldsymbol{s}$ ) is an image, we use convolutional neural networks (CNNs) to represent both actor and critic. Also, instead of using a single Q-value network for the critic, we use two Q-value networks and take their minimum to estimate better the Q-value, as proposed in ref. [Reference Van Hasselt, Guez and Silver38]. To stabilise the learning, we use two more neural networks to represent target Q-values for each critic network, as described in DQN [Reference Mnih, Kavukcuoglu, Silver, Graves, Antonoglou, Wierstra and Riedmiller1]. Therefore, we use 5 CNNs to implement the SAC algorithm.

Figure 2. CNN architecture of policy network. It has three convolutional layers with max-pooling following the first two. The convolutional layers are followed by a four-layer MLP (multilayer perceptron). The output layer is multiheaded, with the number of heads corresponding to the number of tasks. Based on the current task index, only one head is enabled during training and testing. Given an input state (an RGB image), the network produces two 7-dimensional vectors, $\mu$ and $\sigma$ , representing the mean and standard deviation of the stochastic policy modelled by a Gaussian distribution.

Figure 3. The Q-network’s CNN architecture is similar to that of the policy network, with two notable distinctions: firstly, the action vector is incorporated into the first layer of the MLP, and secondly, the network’s output is a scalar Q-value. This network takes state (an RGB image) and action (a 7-dimensional vector) and produces scalar Q-value, that is, $Q(\mathbf{s},\mathbf{a})$ .

Figures 2 and 3 illustrate the architectures of the policy and the Q-value neural networks, respectively. As we parameterise the policy and Q-value using neural networks, $\phi$ represents the set of weights of the policy network. $\theta _1$ , $\theta _2$ , $\hat{\theta }_1$ , and $\hat{\theta }_2$ represent the set of weights of two Q-value networks and two target Q-value networks for the critic, respectively. Since our policy is stochastic, we use tanh-Gaussian policy, as used in ref. [Reference Singh, Yu, Yang, Zhang, Kumar and Levine36]. The policy network takes the state as input and outputs the mean ( $\mu$ ) and standard deviation ( $\sigma$ ) of the Gaussian distribution for each action. Action is then sampled from this distribution and passed through tanh function to bound actions between $(\!-\!1, 1)$ . Target Q-value is defined as

(3) \begin{equation} \hat{Q}_{\hat{\theta }_1,\hat{\theta }_2}(\mathbf{s}_{t+1}, \mathbf{a}_{t+1}) = \\ \mathbf{r}_t + \gamma \mathbb{E}_{(\mathbf{s}_{t+1} \sim \mathcal{D}, \mathbf{a}_{t+1} \sim \pi _\phi (\!\cdot |\mathbf{s}_{t+1}))}[\hat{Q}_{\text{{min}}} - \alpha \text{log}(\pi _\phi (\mathbf{a}_{t+1}|\mathbf{s}_{t+1}))], \end{equation}

where $\hat{Q}_{\text{{min}}}$ represents the minimum Q-value of both target Q-networks and is given by

(4) \begin{equation} \hat{Q}_{\text{{min}}}(\mathbf{s}_{t+1}, \mathbf{a}_{t+1}) = \text{ min}[Q_{\hat{\theta }_1}(\mathbf{s}_{t+1}, \mathbf{a}_{t+1}), Q_{\hat{\theta }_2}(\mathbf{s}_{t+1}, \mathbf{a}_{t+1})] \end{equation}

The target Q-value in (3) is then used to calculate Q-loss for each critic network as

(5) \begin{equation} J_Q({\theta _i}) = \frac{1}{2} \mathbb{E}_{(\mathbf{s}_t, \mathbf{a}_t) \sim \mathcal{D}}[(\hat{Q}_{\hat{\theta }_i,\hat{\theta }_2}(\mathbf{s}_{t+1}, \mathbf{a}_{t+1}) - Q_{\theta _1}(\mathbf{s}_t, \mathbf{a}_t))^2], \end{equation}

where $i\in \{1,2\}$ , $\mathbf{a}_t^\pi$ is the action sampled from policy $\pi _\phi$ for state $\mathbf{s}_t$ and $\mathcal{D}$ represents the current task data. Further, policy-loss for actor-network is defined as

(6) \begin{equation} J_\pi (\phi ) =\mathbb{E}_{(\mathbf{s}_t \sim \mathcal{D}, \mathbf{a}_t \sim \pi _\phi (\!\cdot |\mathbf{s}_t))}[\alpha \text{log}(\pi _\phi (\mathbf{a}_t|\mathbf{s}_t)) -\text{ min}[Q_{\theta _1}(\mathbf{s}_t, \mathbf{a}_t^\pi ), Q_{\theta _2}(\mathbf{s}_t, \mathbf{a}_t^\pi )]] \end{equation}

For offline-RL, we use the non-Lagrange version of the conservative Q-learning (CQL) approach proposed in ref. [Reference Kumar, Zhou, Tucker and Levine37] as it only requires adding a regularisation loss to already well-established continuous RL methods like SAC. Upon adding this CQL-loss to (5), total Q-loss becomes

(7) \begin{equation} J_Q^{\text{total}}({\theta _i}) = J_Q({\theta _i})+\alpha _{\text{cql}} \mathbb{E}_{\mathbf{s}_t \sim \mathcal{D}}[\text{log}\sum \limits _{\mathbf{a}_t} \text{exp}(Q_{\theta _i}(\mathbf{s}_t, \mathbf{a}_t))-\mathbb{E}_{\mathbf{a}_t \sim \mathcal{D}}[Q_{\theta _i}(\mathbf{s}_t, \mathbf{a}_t)]], \end{equation}

where $i\in \{1,2\}$ , $\alpha _{\text{cql}}$ controls the amount of CQL-loss to be added to Q-loss to penalise actions that are too far away from the existing trajectories, thus keeping the policy conservative in the sense of exploration. These losses are then used to update actor and critic networks using the Adam [Reference Kingma and Ba39] optimisation algorithm.

3.2. Applying synaptic intelligence in offline RL

Synaptic intelligence is a regularisation-based algorithm proposed in ref. [Reference Zenke, Poole and Ganguli35] for sequential task learning. It regularises the loss function of a task with a quadratic loss function as defined in (8) to reduce catastrophic forgetting.

(8) \begin{equation} L_\mu = \sum \limits _{k} \Omega ^\mu _k(\tilde{\phi }_k-\phi _k)^2 \end{equation}

Here, $L_\mu$ is the SI loss for the current task being learned with index $\mu$ , $\phi _k$ is $k$ -th weight of the policy network, and $\tilde{\phi }_k$ is the reference weight corresponding to policy network parameters at the end of the previous task. $\Omega ^\mu _k$ is per-parameter regularisation strength. For more details on calculating $\Omega ^\mu _k$ , refer to [Reference Zenke, Poole and Ganguli35]. SI algorithm penalises neural network weights based on their contributions to the change in the overall loss function. Weights that contributed more to the previous tasks are penalised more and thus do not deviate much from their original values, while the other weights help in learning new tasks. SI defines the importance of weights as the sum of the gradients over the training trajectory, as this approximates the contribution to the reduction in the overall loss function. We use a similar approach to apply SI to Offline-RL as presented in ref. [Reference Wołczyk, Zajac, Pascanu, Kuciński and Miłoś31]. Although the authors did not use SI or offline RL, the approach is similar to applying any regularisation-based continual learning method for the actor-critic RL framework. We regularise the actor to reduce forgetting on previous tasks while learning new tasks using offline reinforcement learning. We add quadratic loss as defined in ref. [Reference Zenke, Poole and Ganguli35] to the policy-loss term in the SAC-CQL algorithm. So now overall policy-loss becomes as described in (9)

(9) \begin{equation} J_{\pi }^{\text{total}}(\phi ) = J_\pi (\phi ) + c L_\mu \end{equation}

Here, $c$ is regularisation strength. Another aspect of continual learning is providing the current task index to the neural network. There are many approaches to tackle this problem, from 1-hot encoding to recognising the task from context. We chose the most straightforward option of a multi-head neural network. Each head of the neural network represents a separate task. Therefore, we select the head for a given task.

4.1. Experimental setup

Our experimental setup is based on a simulated environment, Roboverse, used in ref. [Reference Singh, Yu, Yang, Zhang, Kumar and Levine36]. It is a GYM [Reference Brockman, Cheung, Pettersson, Schneider, Schulman, Tang and Zaremba30]-like environment based upon open-source physics simulator py-bullet [Reference Erwin and Yunfei40]. We collected data for three tasks using this simulated environment.

4.1.1. Object space

We define object space as a subset of the robot’s workspace where the task’s target object is to be placed. In our case, it is a rectangular area on the table before the robot. When initialising the task, the target object is randomly placed in the object space. Fig. 4 object space of all three tasks is visible as a rectangle on the table.

Table I. Details of the collected 6 datasets for each task. The number of trajectories decreases as the area of the object space decreases to maintain a consistent object-space density.

4.1.3. Data collection

To examine the impact of the object-space area and object placement density, we collected a dataset for each combination of the object-space area ( $40\,\textrm{cm}^{2}$ , $360\,\textrm{cm}^{2}$ , and $1000\,\textrm{cm}^{2}$ ) and the object placement density (10 and 20 placements per $\textrm{cm}^{2}$ ) resulting in a total of six datasets for each task. Table I shows the quantitative details of 6 datasets for each task. It shows the number of trajectories and images collected for all combinations of object density and size of area. The length of a single episode (single trajectory) is 20; therefore, 20 data points per trajectory are collected. Format of each data point is $\langle \boldsymbol{s}_t, \boldsymbol{a}_t, r_t, \boldsymbol{s}_{t+1}\rangle$ , here $\boldsymbol{s}_t$ is $48\times 48\times 3$ RGB image. Fig. 4 displays trajectories (in green colour) and reward distribution across object space in the dataset for all three tasks, with an object-space area of 360 cm $^2$ and a density of 20 object configurations per cm $^2$ . It can be seen that when the object is placed closer to the robot, the reward is high as the task is completed in a few steps, while it becomes low as the object moves away.

Figure 4. Top row displays sampled trajectories and bottom row displays the scatter plot of the reward distribution for tasks button-press, pick-shed and open drawer with object-space area 1000 $\text{cm}^2$ and density 20 object configurations per $\text{cm}^2$ . The robot’s base is at (0.6m, 0.0m), which is shown as a black semicircle.

Hand-designed policies

We collect data by employing hand-designed policies. The core of the hand-designed policies revolves around action selection, where the policy decides the next action based on the robot’s current state and the task’s requirements. These actions include moving the end-effector, closing/opening the gripper, lifting the end-effector upward, or halting all movement. The policy has full access to simulation and thus can use various parameters required for task completion, which are otherwise unavailable to the RL agent. We use the following hand-designed policies

1. 1. Press-button Policy: The policy first calculates the distance between the gripper and the button. It moves the gripper towards the top of the button until a certain threshold is crossed. Once the gripper is on the top of the button, it is moved down to press it. The state of the button is continuously monitored at every step. The task is considered successful if the button is pressed and a reward of 1 is awarded per step until the end of the episode.

2. 2. Pick-shed Policy: The policy calculates the distance between gripper and object. If the distance is greater than a threshold, the gripper is moved in the direction of the object. If the gripper and object are close enough, the policy gives action to close the gripper. Then, the object is lifted up in the z direction until a certain height threshold.

3. 3. Open Drawer Policy: Similar to the above policy, this policy also calculates the distance between the drawer handle and gripper. The gripper is moved towards the door handle until a certain threshold. Once the gripper is above the handle, the gripper is closed. Finally, the gripper is moved in the direction of opening the drawer. This direction depends on the drawer’s orientation, which is provided in the policy from the simulation.

Our data collection procedure, as defined in algorithm 1, follows the standard agent-environment loop. We initiate an outer loop to iterate over the total number of trajectories (or episodes) $N_T$ , which is determined by considering the area and density of the object space, ensuring a consistent density across different areas. Each episode consists of 20 time steps. During each time step, the hand-designed policy generates an action ( $\boldsymbol{a}_t$ ) based on the current environment state. Then, a Gaussian noise ( $N$ ) is added to the action. The purpose of this noise is twofold: it reduces the policy’s accuracy from 100% to approximately 80% so that we get both successful and failure cases, and it introduces variations in the trajectories. The noisy action is then provided to the simulator, which produces the next state ( $\boldsymbol{s}_{t+1}$ ) and the corresponding reward ( $r_t$ ). We store this information as a typical tuple $\lt \boldsymbol{s}_t, \boldsymbol{a}_t, r_t, \boldsymbol{s}_{t+1}\gt$ , which is commonly utilised in reinforcement learning.

Algorithm 1. Data Collection Procedure.

4.2. Empirical results and analysis

For one sequential learning experiment, we select a sequence of two tasks from the three tasks set, as mentioned in the previous section. This selection yields six possible combinations: button-shed, button-drawer, shed-button, shed-drawer, drawer-shed, and drawer-button. We perform two sets of experiments for each doublet sequence, one with SI regularisation and another without SI regularisation. Each set contains six experiments by varying the area and density of object space. Therefore, in total, we performed 72 experiments of sequential learning. Apart from these 72 experiments, we also trained the agent for single tasks using SAC-CQL for reference baseline performance to evaluate forward transfer. We do behaviour cloning for the initial 5k steps to learn faster as we have a limited compute budget. We use metrics mentioned in ref. [Reference Wołczyk, Zajac, Pascanu, Kuciński and Miłoś31] for evaluating the performance of a continual learning agent. Each task is trained for $\Delta = 100K$ steps. The total number of tasks in a sequence is $N=2$ . Total steps $T = 2 \cdot \Delta$ . The $i\text{-th}$ task is trained from $t \in [(i-1) \cdot \Delta, i \cdot \Delta ]$ .

4.2.1 Task accuracy

We evaluate the agent after every 1000 training steps by sampling ten trajectories from the environment for each task. The agent’s accuracy $p_i(t)$ , for a task $i$ , is defined as the number of successful trajectories out of those ten trials. Fig. 5 shows the accuracy of three experiments corresponding to button-shed, button-drawer, and drawer-button task combinations for the area size of $40cm^2$ with a density of 10 and 20 object configurations per $cm^2$ . The top row represents sequential learning with SI, while the bottom represents sequential learning without SI. SMMMMMMMMI is working better, as evidenced by overlapping Task-1 and Task-2 accuracy. We observed that SI was most helpful in button-shed task doublet due to the overlapping nature of these tasks, as both require reaching the object. This shows the benefit of using SI for overlapping tasks.

Figure 5. Task accuracy for tasks button-shed, button-drawer and drawer-button. The top row is with SI, and the bottom row is without SI. In each plot, the X-axis represents the number of gradient update steps, and the Y-axis represents accuracy.

4.2.2 Forgetting

It measures the decrease in accuracy of the task as we train more tasks and is defined as $F_i:= p_i(i.\Delta )-p_i(T)$ . Here, $p_i(t) \in [0,1]$ is the success rate of task $i$ at time $t$ . Fig. 6 shows the forgetting of Task-1 after training Task-2. We can see that SI performed better or equal in all cases. In fact, in some cases, like button-shed forgetting is negative, which means that the performance of Task-1 improved after training on Task-2. This indicates knowledge transfer from Task-1 to Task-2. This phenomenon is not seen in the case of sequential learning without SI. This indicates that SI helps in reducing catastrophic forgetting. No significant trends are observed in the variation of object-space area, but forgetting increases with increased object-space density. This might be due to the limited computing budget (100K) per task, as tasks with more area size and density would require more training to show good results.

Figure 6. Forgetting matrix (0-button, 1-shed, 2-drawer). The top row is with SI regularisation, and the bottom row is without regularisation. Measuring units for area and density are cm $^2$ and objects/cm $^2$ respectively for every task.

4.2.3 Forward transfer

It measures knowledge transfer by comparing the performance of a given task when trained individually versus learning the task after the network is already trained on previous tasks and is defined as

(10) \begin{equation} FT_i := \frac{\text{AUC}_i - \text{AUC}_i^b}{1- \text{AUC}_i^b}, \end{equation}

where $\text{AUC}_i = \frac{1}{\Delta } \int _{(i-1) \cdot \Delta }^{i \cdot \Delta } p_i(t)\text{d}t$ represents area under the accuracy curve of task $i$ and $\text{AUC}_i^b = \frac{1}{\Delta } \int _{0}^{\Delta } p_i^b(t)\text{d}t$ represents area under curve of the reference baseline task. $p_i^b(t)$ represents reference baseline performance. We use single-task training performance as the reference for Task-2 while evaluating forward transfer. Fig. 7 shows forward transfer for Task-2 after it is trained on Task-1. We observed that in most cases, training without SI gives a better transfer ratio than training with SI. Since reducing catastrophic forgetting is the primary objective of the sequential learning framework, we set a high value of SI regularisation strength. This restricts the movement of weights from the solution of the previous task, which helps to reduce catastrophic forgetting but also hinders the ability to learn new task thus reducing forward transfer. This can also be noticed in Fig. 5, where the accuracy of Task-2 is lower for SI than its non-SI counterpart. This highlights the problem of stability-plasticity; any method that tries to make learning more stable to reduce forgetting inadvertently also restricts the flexibility of the connectionist model to learn a new task.

4.2.4 Training time

We used NVIDIA DGX A100 GPU for training. Training time for one experiment with three sequential tasks on 1 GPU is 18h (6h per task). We used 8 GPUs in parallel for training and testing all 72 experiments, which took approximately 7 days. Apart from these metrics, we observed that the agent requires 14k, 10k, and 16k steps on average to achieve its first success on Task-2 when trained directly, sequentially without SI, and sequentially with SI, respectively. This shows the advantage of the sequential training framework as the agent learns the task faster when trained sequentially (without SI) than when directly training the task, but the agent slows down a little when we add SI to reduce forgetting.

Figure 7. Forward transfer matrix (0-button, 1-shed, 2-drawer). The top row is with SI regularisation, and the bottom row is without regularisation. Measuring units for area and density are cm $^2$ and objects/cm $^2$ respectively for every task.

Figure 8. Accuracy for sequentially learning pick-shed and press-button tasks (area = 360cm $^2$ , density = 20objects/cm $^2$ ). The left column is with SI, and the right column is without SI. In each plot, the X-axis represents the number of gradient update steps, and the Y-axis represents accuracy.

Fig. 8 shows another interesting observation we made in the case of sequential learning of pick-shed and press-button (area = 360cm $^2$ , density = 20objects/cm $^2$ ) tasks. While training for Task-1 (pick shed), the agent showed some success on Task-2 (press button) even before getting any success on Task-1 itself. This might be due to the nature of the tasks, as the trajectory of the press-button task is common for another task. Therefore, the agent tends to acquire knowledge for similar tasks. This may also result from behaviour cloning for the initial 5k steps, where the agent tries to mimic the data collection policy for a few initial training steps. Also, we observed that increasing the object-space area (keeping the density the same) helps in knowledge transfer, which the increase can be seen in average forward transfer with area size.